Fixed locking for non-MVCC snapshots

This reverts commit dbef8d1ad1.

.dockerignore

@@ -0,0 +1,8 @@
+/.git/
+/dist/
+/results/
+/tmp_check/
+/sql/vector--?.?.?.sql
+regression.*
+*.o
+*.so

.editorconfig

@@ -1,6 +1,6 @@
 root = true
 
-[*.{c,h}]
+[*.{c,h,pl,pm,sql}]
 indent_style = tab
 indent_size = tab
 tab_width = 4

.github/workflows/build.yml

@@ -1,39 +1,100 @@
 name: build
 on: [push, pull_request]
 jobs:
-  build:
-    if: "!contains(github.event.head_commit.message, '[skip ci]')"
+  ubuntu:
     runs-on: ${{ matrix.os }}
+    if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest]
-        postgres: [13, 12, 11, 10, 9.6]
         include:
-        - os: macos-latest
-          postgres: 13
+          - postgres: 16
+            os: ubuntu-22.04
+          - postgres: 15
+            os: ubuntu-22.04
+          - postgres: 14
+            os: ubuntu-22.04
+          - postgres: 13
+            os: ubuntu-20.04
+          - postgres: 12
+            os: ubuntu-20.04
+          - postgres: 11
+            os: ubuntu-20.04
     steps:
-    - uses: actions/checkout@v2
-    - uses: ankane/setup-postgres@v1
-      with:
-        postgres-version: ${{ matrix.postgres }}
-    - if: ${{ startsWith(matrix.os, 'ubuntu') }}
-      run: sudo apt-get update && sudo apt-get install postgresql-server-dev-${{ matrix.postgres }} libipc-run-perl
-    - run: make
-    - if: ${{ startsWith(matrix.os, 'ubuntu') }}
-      run: |
-        export PG_CONFIG=`which pg_config`
-        sudo --preserve-env=PG_CONFIG make install
-    - if: ${{ startsWith(matrix.os, 'macos') }}
-      run: make install
-    - run: make installcheck
-    - if: ${{ failure() }}
-      run: cat regression.diffs
-    - if: ${{ startsWith(matrix.os, 'ubuntu') }}
-      run: make prove_installcheck
-    - if: ${{ startsWith(matrix.os, 'macos') }}
-      run: |
-        brew install cpanm && cpanm IPC::Run
-        wget -q https://github.com/postgres/postgres/archive/refs/tags/REL_13_2.tar.gz
-        tar xf REL_13_2.tar.gz
-        make prove_installcheck PROVE=prove PERL5LIB=postgres-REL_13_2/src/test/perl
+      - uses: actions/checkout@v3
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: ${{ matrix.postgres }}
+          dev-files: true
+      - run: make
+        env:
+          PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
+      - run: |
+          export PG_CONFIG=`which pg_config`
+          sudo --preserve-env=PG_CONFIG make install
+      - run: make installcheck
+      - if: ${{ failure() }}
+        run: cat regression.diffs
+      - run: |
+          sudo apt-get update
+          sudo apt-get install libipc-run-perl
+      - run: make prove_installcheck
+  mac:
+    runs-on: macos-latest
+    if: ${{ !startsWith(github.ref_name, 'windows') }}
+    steps:
+      - uses: actions/checkout@v3
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: 14
+      - run: make
+        env:
+          PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter
+      - run: make install
+      - run: make installcheck
+      - if: ${{ failure() }}
+        run: cat regression.diffs
+      - run: |
+          brew install cpanm
+          cpanm --notest IPC::Run
+          wget -q https://github.com/postgres/postgres/archive/refs/tags/REL_14_5.tar.gz
+          tar xf REL_14_5.tar.gz
+      - run: make prove_installcheck PROVE_FLAGS="-I ./postgres-REL_14_5/src/test/perl" PERL5LIB="/Users/runner/perl5/lib/perl5"
+      - run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make
+  windows:
+    runs-on: windows-latest
+    if: ${{ !startsWith(github.ref_name, 'mac') }}
+    steps:
+      - uses: actions/checkout@v3
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: 14
+      - run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat" && ^
+          nmake /NOLOGO /F Makefile.win && ^
+          nmake /NOLOGO /F Makefile.win install && ^
+          nmake /NOLOGO /F Makefile.win installcheck && ^
+          nmake /NOLOGO /F Makefile.win clean && ^
+          nmake /NOLOGO /F Makefile.win uninstall
+        shell: cmd
+  i386:
+    if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
+    runs-on: ubuntu-latest
+    container:
+      image: debian:11
+      options: --platform linux/386
+    steps:
+      - run: apt-get update && apt-get install -y build-essential git libipc-run-perl postgresql-13 postgresql-server-dev-13 sudo
+      - run: service postgresql start
+      - run: |
+          git clone https://github.com/${{ github.repository }}.git pgvector
+          cd pgvector
+          git fetch origin ${{ github.ref }}
+          git reset --hard FETCH_HEAD
+          make
+          make install
+          chown -R postgres .
+          sudo -u postgres make installcheck
+          sudo -u postgres make prove_installcheck
+        env:
+          PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare

.gitignore

@@ -1,5 +1,14 @@
-results
-tmp_check
+/dist/
+/log/
+/results/
+/tmp_check/
+/sql/vector--?.?.?.sql
 regression.*
 *.o
 *.so
+*.bc
+*.dll
+*.dylib
+*.obj
+*.lib
+*.exp

CHANGELOG.md

@@ -1,3 +1,138 @@
+## 0.5.1 (unreleased)
+
+- Improved performance of index scans for IVFFlat after updates and deletes
+- Fixed locking for index scans for HNSW
+
+## 0.5.0 (2023-08-28)
+
+- Added HNSW index type
+- Added support for parallel index builds for IVFFlat
+- Added `l1_distance` function
+- Added element-wise multiplication for vectors
+- Added `sum` aggregate
+- Improved performance of distance functions
+- Fixed out of range results for cosine distance
+- Fixed results for NULL and NaN distances for IVFFlat
+
+## 0.4.4 (2023-06-12)
+
+- Improved error message for malformed vector literal
+- Fixed segmentation fault with text input
+- Fixed consecutive delimiters with text input
+
+## 0.4.3 (2023-06-10)
+
+- Improved cost estimation
+- Improved support for spaces with text input
+- Fixed infinite and NaN values with binary input
+- Fixed infinite values with vector addition and subtraction
+- Fixed infinite values with list centers
+- Fixed compilation error when `float8` is pass by reference
+- Fixed compilation error on PowerPC
+- Fixed segmentation fault with index creation on i386
+
+## 0.4.2 (2023-05-13)
+
+- Added notice when index created with little data
+- Fixed dimensions check for some direct function calls
+- Fixed installation error with Postgres 12.0-12.2
+
+## 0.4.1 (2023-03-21)
+
+- Improved performance of cosine distance
+- Fixed index scan count
+
+## 0.4.0 (2023-01-11)
+
+If upgrading with Postgres < 13, see [this note](https://github.com/pgvector/pgvector#040).
+
+- Changed text representation for vector elements to match `real`
+- Changed storage for vector from `plain` to `extended`
+- Increased max dimensions for vector from 1024 to 16000
+- Increased max dimensions for index from 1024 to 2000
+- Improved accuracy of text parsing for certain inputs
+- Added `avg` aggregate for vector
+- Added experimental support for Windows
+- Dropped support for Postgres 10
+
+## 0.3.2 (2022-11-22)
+
+- Fixed `invalid memory alloc request size` error
+
+## 0.3.1 (2022-11-02)
+
+If upgrading from 0.2.7 or 0.3.0, [recreate](https://github.com/pgvector/pgvector#031) all `ivfflat` indexes after upgrading to ensure all data is indexed.
+
+- Fixed issue with inserts silently corrupting `ivfflat` indexes (introduced in 0.2.7)
+- Fixed segmentation fault with index creation when lists > 6500
+
+## 0.3.0 (2022-10-15)
+
+- Added support for Postgres 15
+- Dropped support for Postgres 9.6
+
+## 0.2.7 (2022-07-31)
+
+- Fixed `unexpected data beyond EOF` error
+
+## 0.2.6 (2022-05-22)
+
+- Improved performance of index creation for Postgres < 12
+
+## 0.2.5 (2022-02-11)
+
+- Reduced memory usage during index creation
+- Fixed index creation exceeding `maintenance_work_mem`
+- Fixed error with index creation when lists > 1600
+
+## 0.2.4 (2022-02-06)
+
+- Added support for parallel vacuum
+- Fixed issue with index not reusing space
+
+## 0.2.3 (2022-01-30)
+
+- Added indexing progress for Postgres 12+
+- Improved interrupt handling during index creation
+
+## 0.2.2 (2022-01-15)
+
+- Fixed compilation error on Mac ARM
+
+## 0.2.1 (2022-01-02)
+
+- Fixed `operator is not unique` error
+
+## 0.2.0 (2021-10-03)
+
+- Added support for Postgres 14
+
+## 0.1.8 (2021-09-07)
+
+- Added cast for `vector` to `real[]`
+
+## 0.1.7 (2021-06-13)
+
+- Added cast for `numeric[]` to `vector`
+
+## 0.1.6 (2021-06-09)
+
+- Fixed segmentation fault with `COUNT`
+
+## 0.1.5 (2021-05-25)
+
+- Reduced memory usage during index creation
+
+## 0.1.4 (2021-05-09)
+
+- Fixed kmeans for inner product
+- Fixed multiple definition error with GCC 10
+
+## 0.1.3 (2021-05-06)
+
+- Added Dockerfile
+- Fixed version
+
 ## 0.1.2 (2021-04-26)
 
 - Vectorized distance calculations

Dockerfile

@@ -0,0 +1,20 @@
+ARG PG_MAJOR=15
+FROM postgres:$PG_MAJOR
+ARG PG_MAJOR
+
+COPY . /tmp/pgvector
+
+RUN apt-get update && \
+		apt-mark hold locales && \
+		apt-get install -y --no-install-recommends build-essential postgresql-server-dev-$PG_MAJOR && \
+		cd /tmp/pgvector && \
+		make clean && \
+		make OPTFLAGS="" && \
+		make install && \
+		mkdir /usr/share/doc/pgvector && \
+		cp LICENSE README.md /usr/share/doc/pgvector && \
+		rm -r /tmp/pgvector && \
+		apt-get remove -y build-essential postgresql-server-dev-$PG_MAJOR && \
+		apt-get autoremove -y && \
+		apt-mark unhold locales && \
+		rm -rf /var/lib/apt/lists/*

LICENSE

@@ -1,4 +1,4 @@
-Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
 
 Portions Copyright (c) 1994, The Regents of the University of California
 

META.json

@@ -0,0 +1,49 @@
+{
+	"name": "vector",
+	"abstract": "Open-source vector similarity search for Postgres",
+	"description": "Supports L2 distance, inner product, and cosine distance",
+	"version": "0.5.0",
+	"maintainer": [
+		"Andrew Kane <andrew@ankane.org>"
+	],
+	"license": {
+		"PostgreSQL": "http://www.postgresql.org/about/licence"
+	},
+	"prereqs": {
+		"runtime": {
+			"requires": {
+				"PostgreSQL": "11.0.0"
+			}
+		}
+	},
+	"provides": {
+		"vector": {
+			"file": "sql/vector.sql",
+			"docfile": "README.md",
+			"version": "0.5.0",
+			"abstract": "Open-source vector similarity search for Postgres"
+		}
+	},
+	"resources": {
+		"homepage": "https://github.com/pgvector/pgvector",
+		"bugtracker": {
+			"web": "https://github.com/pgvector/pgvector/issues"
+		},
+		"repository": {
+			"url":  "https://github.com/pgvector/pgvector.git",
+			"web":  "https://github.com/pgvector/pgvector",
+			"type": "git"
+		}
+	},
+	"generated_by": "Andrew Kane",
+	"meta-spec": {
+		"version": "1.0.0",
+		"url": "http://pgxn.org/meta/spec.txt"
+	},
+	"tags": [
+		"vectors",
+		"datatype",
+		"nearest neighbor search",
+		"approximate nearest neighbors"
+	]
+}

Makefile

@@ -1,16 +1,34 @@
 EXTENSION = vector
-DATA = vector--0.1.1.sql vector--0.1.0--0.1.1.sql
+EXTVERSION = 0.5.0
+
 MODULE_big = vector
-OBJS = src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
+DATA = $(wildcard sql/*--*.sql)
+OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
+HEADERS = src/vector.h
 
-TESTS = $(wildcard sql/*.sql)
+TESTS = $(wildcard test/sql/*.sql)
+REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
+REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
 
-REGRESS = btree cast copy functions ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_unlogged vector
+OPTFLAGS = -march=native
+
+# Mac ARM doesn't support -march=native
+ifeq ($(shell uname -s), Darwin)
+	ifeq ($(shell uname -p), arm)
+		# no difference with -march=armv8.5-a
+		OPTFLAGS =
+	endif
+endif
+
+# PowerPC doesn't support -march=native
+ifneq ($(filter ppc64%, $(shell uname -m)), )
+	OPTFLAGS =
+endif
 
 # For auto-vectorization:
 # - GCC (needs -ftree-vectorize OR -O3) - https://gcc.gnu.org/projects/tree-ssa/vectorization.html
 # - Clang (could use pragma instead) - https://llvm.org/docs/Vectorizers.html
-PG_CFLAGS = -march=native -ftree-vectorize -fassociative-math -fno-signed-zeros -fno-trapping-math
+PG_CFLAGS += $(OPTFLAGS) -ftree-vectorize -fassociative-math -fno-signed-zeros -fno-trapping-math
 
 # Debug GCC auto-vectorization
 # PG_CFLAGS += -fopt-info-vec
@@ -18,10 +36,42 @@ PG_CFLAGS = -march=native -ftree-vectorize -fassociative-math -fno-signed-zeros
 # Debug Clang auto-vectorization
 # PG_CFLAGS += -Rpass=loop-vectorize -Rpass-analysis=loop-vectorize
 
+all: sql/$(EXTENSION)--$(EXTVERSION).sql
+
+sql/$(EXTENSION)--$(EXTVERSION).sql: sql/$(EXTENSION).sql
+	cp $< $@
+
+EXTRA_CLEAN = sql/$(EXTENSION)--$(EXTVERSION).sql
+
 PG_CONFIG ?= pg_config
 PGXS := $(shell $(PG_CONFIG) --pgxs)
 include $(PGXS)
 
+# for Mac
+ifeq ($(PROVE),)
+	PROVE = prove
+endif
+
+# for Postgres 15
+PROVE_FLAGS += -I ./test/perl
+
 prove_installcheck:
 	rm -rf $(CURDIR)/tmp_check
-	cd $(srcdir) && TESTDIR='$(CURDIR)' PATH="$(bindir):$$PATH" PGPORT='6$(DEF_PGPORT)' PG_REGRESS='$(top_builddir)/src/test/regress/pg_regress' $(PROVE) $(PG_PROVE_FLAGS) $(PROVE_FLAGS) $(if $(PROVE_TESTS),$(PROVE_TESTS),t/*.pl)
+	cd $(srcdir) && TESTDIR='$(CURDIR)' PATH="$(bindir):$$PATH" PGPORT='6$(DEF_PGPORT)' PG_REGRESS='$(top_builddir)/src/test/regress/pg_regress' $(PROVE) $(PG_PROVE_FLAGS) $(PROVE_FLAGS) $(if $(PROVE_TESTS),$(PROVE_TESTS),test/t/*.pl)
+
+.PHONY: dist
+
+dist:
+	mkdir -p dist
+	git archive --format zip --prefix=$(EXTENSION)-$(EXTVERSION)/ --output dist/$(EXTENSION)-$(EXTVERSION).zip master
+
+.PHONY: docker
+
+docker:
+	docker build --pull --no-cache --platform linux/amd64 -t ankane/pgvector:latest .
+
+.PHONY: docker-release
+
+docker-release:
+	docker buildx build --push --pull --no-cache --platform linux/amd64,linux/arm64 -t ankane/pgvector:latest .
+	docker buildx build --push --platform linux/amd64,linux/arm64 -t ankane/pgvector:v$(EXTVERSION) .

Makefile.win

@@ -0,0 +1,75 @@
+EXTENSION = vector
+EXTVERSION = 0.5.0
+
+OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
+HEADERS = src\vector.h
+
+REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
+REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
+
+# For /arch flags
+# https://learn.microsoft.com/en-us/cpp/build/reference/arch-minimum-cpu-architecture
+OPTFLAGS =
+
+# For auto-vectorization:
+# - MSVC (needs /O2 /fp:fast) - https://learn.microsoft.com/en-us/cpp/parallel/auto-parallelization-and-auto-vectorization?#auto-vectorizer
+PG_CFLAGS = $(PG_CFLAGS) $(OPTFLAGS) /O2 /fp:fast
+
+# Debug MSVC auto-vectorization
+# https://learn.microsoft.com/en-us/cpp/error-messages/tool-errors/vectorizer-and-parallelizer-messages
+# PG_CFLAGS = $(PG_CFLAGS) /Qvec-report:2
+
+all: sql\$(EXTENSION)--$(EXTVERSION).sql
+
+sql\$(EXTENSION)--$(EXTVERSION).sql: sql\$(EXTENSION).sql
+	copy sql\$(EXTENSION).sql $@
+
+# TODO use pg_config
+!ifndef PGROOT
+!error PGROOT is not set
+!endif
+BINDIR = $(PGROOT)\bin
+INCLUDEDIR = $(PGROOT)\include
+INCLUDEDIR_SERVER = $(PGROOT)\include\server
+LIBDIR = $(PGROOT)\lib
+PKGLIBDIR = $(PGROOT)\lib
+SHAREDIR = $(PGROOT)\share
+
+CFLAGS = /nologo /I"$(INCLUDEDIR_SERVER)\port\win32_msvc" /I"$(INCLUDEDIR_SERVER)\port\win32" /I"$(INCLUDEDIR_SERVER)" /I"$(INCLUDEDIR)"
+
+CFLAGS = $(CFLAGS) $(PG_CFLAGS)
+
+SHLIB = $(EXTENSION).dll
+
+LIBS = "$(LIBDIR)\postgres.lib"
+
+.c.obj:
+	$(CC) $(CFLAGS) /c $< /Fo$@
+
+$(SHLIB): $(OBJS)
+	$(CC) $(CFLAGS) $(OBJS) $(LIBS) /link /DLL /OUT:$(SHLIB)
+
+all: $(SHLIB)
+
+install:
+	copy $(SHLIB) "$(PKGLIBDIR)"
+	copy $(EXTENSION).control "$(SHAREDIR)\extension"
+	copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
+	mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+	copy $(HEADERS) "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+
+installcheck:
+	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
+
+uninstall:
+	del /f "$(PKGLIBDIR)\$(SHLIB)"
+	del /f "$(SHAREDIR)\extension\$(EXTENSION).control"
+	del /f "$(SHAREDIR)\extension\$(EXTENSION)--*.sql"
+	del /f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)\*.h"
+	rmdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+
+clean:
+	del /f $(SHLIB) $(EXTENSION).lib $(EXTENSION).exp
+	del /f $(OBJS)
+	del /f sql\$(EXTENSION)--$(EXTVERSION).sql
+	del /f /s /q results regression.diffs regression.out tmp_check tmp_check_iso log output_iso

README.md

@@ -2,148 +2,635 @@
 
 Open-source vector similarity search for Postgres
 
-```sql
-CREATE TABLE table (column vector(3));
-CREATE INDEX ON table USING ivfflat (column);
-SELECT * FROM table ORDER BY column <-> '[1,2,3]' LIMIT 5;
-```
+Store your vectors with the rest of your data. Supports:
 
-Supports L2 distance, inner product, and cosine distance
+- exact and approximate nearest neighbor search
+- L2 distance, inner product, and cosine distance
+- any [language](#languages) with a Postgres client
 
-[![Build Status](https://github.com/ankane/pgvector/workflows/build/badge.svg?branch=master)](https://github.com/ankane/pgvector/actions)
+Plus [ACID](https://en.wikipedia.org/wiki/ACID) compliance, point-in-time recovery, JOINs, and all of the other [great features](https://www.postgresql.org/about/) of Postgres
+
+[![Build Status](https://github.com/pgvector/pgvector/workflows/build/badge.svg?branch=master)](https://github.com/pgvector/pgvector/actions)
 
 ## Installation
 
-Compile and install the extension (supports Postgres 9.6+)
+Compile and install the extension (supports Postgres 11+)
 
 ```sh
-git clone --branch v0.1.2 https://github.com/ankane/pgvector.git
+cd /tmp
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
 ```
 
-Then load it in databases where you want to use it
+See the [installation notes](#installation-notes) if you run into issues
 
-```sql
-CREATE EXTENSION vector;
-```
+You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres)
 
 ## Getting Started
 
-Create a vector column with 3 dimensions (replace `table` and `column` with non-reserved names)
+Enable the extension (do this once in each database where you want to use it)
 
-```sql
-CREATE TABLE table (column vector(3));
+```tsql
+CREATE EXTENSION vector;
 ```
 
-Insert values
+Create a vector column with 3 dimensions
 
 ```sql
-INSERT INTO table VALUES ('[1,2,3]'), ('[4,5,6]');
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
 ```
 
-Get the nearest neighbor by L2 distance
+Insert vectors
 
 ```sql
-SELECT * FROM table ORDER BY column <-> '[3,1,2]' LIMIT 1;
+INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
+```
+
+Get the nearest neighbors by L2 distance
+
+```sql
+SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
 ```
 
 Also supports inner product (`<#>`) and cosine distance (`<=>`)
 
 Note: `<#>` returns the negative inner product since Postgres only supports `ASC` order index scans on operators
 
+## Storing
+
+Create a new table with a vector column
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
+```
+
+Or add a vector column to an existing table
+
+```sql
+ALTER TABLE items ADD COLUMN embedding vector(3);
+```
+
+Insert vectors
+
+```sql
+INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
+```
+
+Upsert vectors
+
+```sql
+INSERT INTO items (id, embedding) VALUES (1, '[1,2,3]'), (2, '[4,5,6]')
+    ON CONFLICT (id) DO UPDATE SET embedding = EXCLUDED.embedding;
+```
+
+Update vectors
+
+```sql
+UPDATE items SET embedding = '[1,2,3]' WHERE id = 1;
+```
+
+Delete vectors
+
+```sql
+DELETE FROM items WHERE id = 1;
+```
+
+## Querying
+
+Get the nearest neighbors to a vector
+
+```sql
+SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
+```
+
+Get the nearest neighbors to a row
+
+```sql
+SELECT * FROM items WHERE id != 1 ORDER BY embedding <-> (SELECT embedding FROM items WHERE id = 1) LIMIT 5;
+```
+
+Get rows within a certain distance
+
+```sql
+SELECT * FROM items WHERE embedding <-> '[3,1,2]' < 5;
+```
+
+Note: Combine with `ORDER BY` and `LIMIT` to use an index
+
+#### Distances
+
+Get the distance
+
+```sql
+SELECT embedding <-> '[3,1,2]' AS distance FROM items;
+```
+
+For inner product, multiply by -1 (since `<#>` returns the negative inner product)
+
+```tsql
+SELECT (embedding <#> '[3,1,2]') * -1 AS inner_product FROM items;
+```
+
+For cosine similarity, use 1 - cosine distance
+
+```sql
+SELECT 1 - (embedding <=> '[3,1,2]') AS cosine_similarity FROM items;
+```
+
+#### Aggregates
+
+Average vectors
+
+```sql
+SELECT AVG(embedding) FROM items;
+```
+
+Average groups of vectors
+
+```sql
+SELECT category_id, AVG(embedding) FROM items GROUP BY category_id;
+```
+
 ## Indexing
 
-Speed up queries with an approximate index. Add an index for each distance function you want to use.
+By default, pgvector performs exact nearest neighbor search, which provides perfect recall.
+
+You can add an index to use approximate nearest neighbor search, which trades some recall for speed. Unlike typical indexes, you will see different results for queries after adding an approximate index.
+
+Supported index types are:
+
+- [IVFFlat](#ivfflat)
+- [HNSW](#hnsw) - added in 0.5.0
+
+## IVFFlat
+
+An IVFFlat index divides vectors into lists, and then searches a subset of those lists that are closest to the query vector. It has faster build times and uses less memory than HNSW, but has lower query performance (in terms of speed-recall tradeoff).
+
+Three keys to achieving good recall are:
+
+1. Create the index *after* the table has some data
+2. Choose an appropriate number of lists - a good place to start is `rows / 1000` for up to 1M rows and `sqrt(rows)` for over 1M rows
+3. When querying, specify an appropriate number of [probes](#query-options) (higher is better for recall, lower is better for speed) - a good place to start is `sqrt(lists)`
+
+Add an index for each distance function you want to use.
 
 L2 distance
 
 ```sql
-CREATE INDEX ON table USING ivfflat (column);
+CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100);
 ```
 
 Inner product
 
 ```sql
-CREATE INDEX ON table USING ivfflat (column vector_ip_ops);
+CREATE INDEX ON items USING ivfflat (embedding vector_ip_ops) WITH (lists = 100);
 ```
 
 Cosine distance
 
 ```sql
-CREATE INDEX ON table USING ivfflat (column vector_cosine_ops);
+CREATE INDEX ON items USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);
 ```
 
-Indexes should be created after the table has data for optimal clustering. Also, unlike typical indexes which only affect performance, you may see different results for queries after adding an approximate index.
-
-### Index Options
-
-Specify the number of inverted lists (100 by default)
-
-```sql
-CREATE INDEX ON table USING ivfflat (column) WITH (lists = 100);
-```
+Vectors with up to 2,000 dimensions can be indexed.
 
 ### Query Options
 
 Specify the number of probes (1 by default)
 
 ```sql
-SET ivfflat.probes = 1;
+SET ivfflat.probes = 10;
 ```
 
-A higher value improves recall at the cost of speed.
+A higher value provides better recall at the cost of speed, and it can be set to the number of lists for exact nearest neighbor search (at which point the planner won’t use the index)
 
 Use `SET LOCAL` inside a transaction to set it for a single query
 
 ```sql
 BEGIN;
-SET LOCAL ivfflat.probes = 1;
+SET LOCAL ivfflat.probes = 10;
 SELECT ...
 COMMIT;
 ```
 
+## HNSW
+
+An HNSW index creates a multilayer graph. It has slower build times and uses more memory than IVFFlat, but has better query performance (in terms of speed-recall tradeoff). There’s no training step like IVFFlat, so the index can be created without any data in the table.
+
+Add an index for each distance function you want to use.
+
+L2 distance
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
+```
+
+Inner product
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
+```
+
+Cosine distance
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
+```
+
+Vectors with up to 2,000 dimensions can be indexed.
+
+### Index Options
+
+Specify HNSW parameters
+
+- `m` - the max number of connections per layer (16 by default)
+- `ef_construction` - the size of the dynamic candidate list for constructing the graph (64 by default)
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WITH (m = 16, ef_construction = 64);
+```
+
+### Query Options
+
+Specify the size of the dynamic candidate list for search (40 by default)
+
+```sql
+SET hnsw.ef_search = 100;
+```
+
+A higher value provides better recall at the cost of speed.
+
+Use `SET LOCAL` inside a transaction to set it for a single query
+
+```sql
+BEGIN;
+SET LOCAL hnsw.ef_search = 100;
+SELECT ...
+COMMIT;
+```
+
+## Indexing Progress
+
+Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
+
+```sql
+SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
+```
+
+The phases are:
+
+1. `initializing`
+2. `performing k-means` - IVFFlat only
+3. `assigning tuples` - IVFFlat only
+4. `loading tuples`
+
+Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
+
+## Filtering
+
+There are a few ways to index nearest neighbor queries with a `WHERE` clause
+
+```sql
+SELECT * FROM items WHERE category_id = 123 ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
+```
+
+Create an index on one [or more](https://www.postgresql.org/docs/current/indexes-multicolumn.html) of the `WHERE` columns for exact search
+
+```sql
+CREATE INDEX ON items (category_id);
+```
+
+Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search
+
+```sql
+CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)
+    WHERE (category_id = 123);
+```
+
+Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html) for approximate search on many different values of the `WHERE` columns
+
+```sql
+CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
+```
+
+## Hybrid Search
+
+Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search ([Python example](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py)).
+
+```sql
+SELECT id, content FROM items, plainto_tsquery('hello search') query
+    WHERE textsearch @@ query ORDER BY ts_rank_cd(textsearch, query) DESC LIMIT 5;
+```
+
+## Performance
+
+Use `EXPLAIN ANALYZE` to debug performance.
+
+```sql
+EXPLAIN ANALYZE SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
+```
+
+### Exact Search
+
+To speed up queries without an index, increase `max_parallel_workers_per_gather`.
+
+```sql
+SET max_parallel_workers_per_gather = 4;
+```
+
+If vectors are normalized to length 1 (like [OpenAI embeddings](https://platform.openai.com/docs/guides/embeddings/which-distance-function-should-i-use)), use inner product for best performance.
+
+```tsql
+SELECT * FROM items ORDER BY embedding <#> '[3,1,2]' LIMIT 5;
+```
+
+### Approximate Search
+
+To speed up queries with an IVFFlat index, increase the number of inverted lists (at the expense of recall).
+
+```sql
+CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 1000);
+```
+
+## Languages
+
+Use pgvector from any language with a Postgres client. You can even generate and store vectors in one language and query them in another.
+
+Language | Libraries / Examples
+--- | ---
+C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
+C# | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
+Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
+Dart | [pgvector-dart](https://github.com/pgvector/pgvector-dart)
+Elixir | [pgvector-elixir](https://github.com/pgvector/pgvector-elixir)
+Go | [pgvector-go](https://github.com/pgvector/pgvector-go)
+Haskell | [pgvector-haskell](https://github.com/pgvector/pgvector-haskell)
+Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
+Julia | [pgvector-julia](https://github.com/pgvector/pgvector-julia)
+Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
+Node.js | [pgvector-node](https://github.com/pgvector/pgvector-node)
+Perl | [pgvector-perl](https://github.com/pgvector/pgvector-perl)
+PHP | [pgvector-php](https://github.com/pgvector/pgvector-php)
+Python | [pgvector-python](https://github.com/pgvector/pgvector-python)
+R | [pgvector-r](https://github.com/pgvector/pgvector-r)
+Ruby | [pgvector-ruby](https://github.com/pgvector/pgvector-ruby), [Neighbor](https://github.com/ankane/neighbor)
+Rust | [pgvector-rust](https://github.com/pgvector/pgvector-rust)
+Swift | [pgvector-swift](https://github.com/pgvector/pgvector-swift)
+
+## Frequently Asked Questions
+
+#### How many vectors can be stored in a single table?
+
+A non-partitioned table has a limit of 32 TB by default in Postgres. A partitioned table can have thousands of partitions of that size.
+
+#### Is replication supported?
+
+Yes, pgvector uses the write-ahead log (WAL), which allows for replication and point-in-time recovery.
+
+#### What if I want to index vectors with more than 2,000 dimensions?
+
+You’ll need to use [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction) at the moment.
+
+## Troubleshooting
+
+#### Why isn’t a query using an index?
+
+The cost estimation in pgvector < 0.4.3 does not always work well with the planner. You can encourage the planner to use an index for a query with:
+
+```sql
+BEGIN;
+SET LOCAL enable_seqscan = off;
+SELECT ...
+COMMIT;
+```
+
+#### Why isn’t a query using a parallel table scan?
+
+The planner doesn’t consider [out-of-line storage](https://www.postgresql.org/docs/current/storage-toast.html) in cost estimates, which can make a serial scan look cheaper. You can reduce the cost of a parallel scan for a query with:
+
+```sql
+BEGIN;
+SET LOCAL min_parallel_table_scan_size = 1;
+SET LOCAL parallel_setup_cost = 1;
+SELECT ...
+COMMIT;
+```
+
+or choose to store vectors inline:
+
+```sql
+ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
+```
+
+#### Why are there less results for a query after adding an IVFFlat index?
+
+The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
+
+```sql
+DROP INDEX index_name;
+```
+
 ## Reference
 
 ### Vector Type
 
-Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a float, and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 1024 dimensions.
+Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a single precision floating-point number (like the `real` type in Postgres), and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 16,000 dimensions.
 
 ### Vector Operators
 
-Operator | Description
---- | ---
-\+ | element-wise addition
-\- | element-wise subtraction
-<-> | Euclidean distance
-<#> | negative inner product
-<=> | cosine distance
+Operator | Description | Added
+--- | --- | ---
+\+ | element-wise addition |
+\- | element-wise subtraction |
+\* | element-wise multiplication | 0.5.0
+<-> | Euclidean distance |
+<#> | negative inner product |
+<=> | cosine distance |
 
 ### Vector Functions
 
-Function | Description
---- | ---
-cosine_distance(vector, vector) | cosine distance
-inner_product(vector, vector) | inner product
-l2_distance(vector, vector) | Euclidean distance
-vector_dims(vector) | number of dimensions
-vector_norm(vector) | Euclidean norm
+Function | Description | Added
+--- | --- | ---
+cosine_distance(vector, vector) → double precision | cosine distance |
+inner_product(vector, vector) → double precision | inner product |
+l2_distance(vector, vector) → double precision | Euclidean distance |
+l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
+vector_dims(vector) → integer | number of dimensions |
+vector_norm(vector) → double precision | Euclidean norm |
 
-## Libraries
+### Aggregate Functions
 
-Libraries that use pgvector:
+Function | Description | Added
+--- | --- | ---
+avg(vector) → vector | average |
+sum(vector) → vector | sum | 0.5.0
 
-- [Neighbor](https://github.com/ankane/neighbor) (Ruby)
+## Installation Notes
+
+### Postgres Location
+
+If your machine has multiple Postgres installations, specify the path to [pg_config](https://www.postgresql.org/docs/current/app-pgconfig.html) with:
+
+```sh
+export PG_CONFIG=/Applications/Postgres.app/Contents/Versions/latest/bin/pg_config
+```
+
+Then re-run the installation instructions (run `make clean` before `make` if needed). If `sudo` is needed for `make install`, use:
+
+```sh
+sudo --preserve-env=PG_CONFIG make install
+```
+
+### Missing Header
+
+If compilation fails with `fatal error: postgres.h: No such file or directory`, make sure Postgres development files are installed on the server.
+
+For Ubuntu and Debian, use:
+
+```sh
+sudo apt install postgresql-server-dev-15
+```
+
+Note: Replace `15` with your Postgres server version
+
+### Windows
+
+Support for Windows is currently experimental. Ensure [C++ support in Visual Studio](https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line?view=msvc-170#download-and-install-the-tools) is installed, and run:
+
+```cmd
+call "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"
+```
+
+Note: The exact path will vary depending on your Visual Studio version and edition
+
+Then use `nmake` to build:
+
+```cmd
+set "PGROOT=C:\Program Files\PostgreSQL\15"
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+cd pgvector
+nmake /F Makefile.win
+nmake /F Makefile.win install
+```
+
+## Additional Installation Methods
+
+### Docker
+
+Get the [Docker image](https://hub.docker.com/r/ankane/pgvector) with:
+
+```sh
+docker pull ankane/pgvector
+```
+
+This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (run it the same way).
+
+You can also build the image manually:
+
+```sh
+git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+cd pgvector
+docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
+```
+
+### Homebrew
+
+With Homebrew Postgres, you can use:
+
+```sh
+brew install pgvector
+```
+
+Note: This only adds it to the `postgresql@14` formula
+
+### PGXN
+
+Install from the [PostgreSQL Extension Network](https://pgxn.org/dist/vector) with:
+
+```sh
+pgxn install vector
+```
+
+### APT
+
+Debian and Ubuntu packages are available from the [PostgreSQL APT Repository](https://wiki.postgresql.org/wiki/Apt). Follow the [setup instructions](https://wiki.postgresql.org/wiki/Apt#Quickstart) and run:
+
+```sh
+sudo apt install postgresql-15-pgvector
+```
+
+Note: Replace `15` with your Postgres server version
+
+### Yum
+
+RPM packages are available from the [PostgreSQL Yum Repository](https://yum.postgresql.org/). Follow the [setup instructions](https://www.postgresql.org/download/linux/redhat/) for your distribution and run:
+
+```sh
+sudo yum install pgvector_15
+# or
+sudo dnf install pgvector_15
+```
+
+Note: Replace `15` with your Postgres server version
+
+### conda-forge
+
+With Conda Postgres, install from [conda-forge](https://anaconda.org/conda-forge/pgvector) with:
+
+```sh
+conda install -c conda-forge pgvector
+```
+
+This method is [community-maintained](https://github.com/conda-forge/pgvector-feedstock) by [@mmcauliffe](https://github.com/mmcauliffe)
+
+### Postgres.app
+
+Download the [latest release](https://postgresapp.com/downloads.html) with Postgres 15+.
+
+## Hosted Postgres
+
+pgvector is available on [these providers](https://github.com/pgvector/pgvector/issues/54).
 
 ## Upgrading
 
-### 0.1.1
-
-Compile and install the latest version and run:
+Install the latest version. Then in each database you want to upgrade, run:
 
 ```sql
-ALTER EXTENSION vector UPDATE TO '0.1.1';
+ALTER EXTENSION vector UPDATE;
+```
+
+You can check the version in the current database with:
+
+```sql
+SELECT extversion FROM pg_extension WHERE extname = 'vector';
+```
+
+## Upgrade Notes
+
+### 0.4.0
+
+If upgrading with Postgres < 13, remove this line from `sql/vector--0.3.2--0.4.0.sql`:
+
+```sql
+ALTER TYPE vector SET (STORAGE = extended);
+```
+
+Then run `make install` and `ALTER EXTENSION vector UPDATE;`.
+
+### 0.3.1
+
+If upgrading from 0.2.7 or 0.3.0, recreate all `ivfflat` indexes after upgrading to ensure all data is indexed.
+
+```sql
+-- Postgres 12+
+REINDEX INDEX CONCURRENTLY index_name;
+
+-- Postgres < 12
+CREATE INDEX CONCURRENTLY temp_name ON table USING ivfflat (column opclass);
+DROP INDEX CONCURRENTLY index_name;
+ALTER INDEX temp_name RENAME TO index_name;
 ```
 
 ## Thanks
@@ -152,27 +639,28 @@ Thanks to:
 
 - [PASE: PostgreSQL Ultra-High-Dimensional Approximate Nearest Neighbor Search Extension](https://dl.acm.org/doi/pdf/10.1145/3318464.3386131)
 - [Faiss: A Library for Efficient Similarity Search and Clustering of Dense Vectors](https://github.com/facebookresearch/faiss)
-- [Using the Triangle Inequality to Accelerate k-means](https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf)
+- [Using the Triangle Inequality to Accelerate k-means](https://cdn.aaai.org/ICML/2003/ICML03-022.pdf)
 - [k-means++: The Advantage of Careful Seeding](https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf)
 - [Concept Decompositions for Large Sparse Text Data using Clustering](https://www.cs.utexas.edu/users/inderjit/public_papers/concept_mlj.pdf)
+- [Efficient and Robust Approximate Nearest Neighbor Search using Hierarchical Navigable Small World Graphs](https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf)
 
 ## History
 
-View the [changelog](https://github.com/ankane/pgvector/blob/master/CHANGELOG.md)
+View the [changelog](https://github.com/pgvector/pgvector/blob/master/CHANGELOG.md)
 
 ## Contributing
 
 Everyone is encouraged to help improve this project. Here are a few ways you can help:
 
-- [Report bugs](https://github.com/ankane/pgvector/issues)
-- Fix bugs and [submit pull requests](https://github.com/ankane/pgvector/pulls)
+- [Report bugs](https://github.com/pgvector/pgvector/issues)
+- Fix bugs and [submit pull requests](https://github.com/pgvector/pgvector/pulls)
 - Write, clarify, or fix documentation
 - Suggest or add new features
 
 To get started with development:
 
 ```sh
-git clone https://github.com/ankane/pgvector.git
+git clone https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install
@@ -188,18 +676,18 @@ make prove_installcheck  # TAP tests
 To run single tests:
 
 ```sh
-make installcheck REGRESS=vector                  # regression test
-make prove_installcheck PROVE_TESTS=t/001_wal.pl  # TAP test
+make installcheck REGRESS=functions                    # regression test
+make prove_installcheck PROVE_TESTS=test/t/001_wal.pl  # TAP test
 ```
 
-Directories
+To enable benchmarking:
 
-- `expected` - expected output for regression tests
-- `sql` - regression tests
-- `t` - TAP tests
+```sh
+make clean && PG_CFLAGS=-DIVFFLAT_BENCH make && make install
+```
 
 Resources for contributors
 
 - [Extension Building Infrastructure](https://www.postgresql.org/docs/current/extend-pgxs.html)
 - [Index Access Method Interface Definition](https://www.postgresql.org/docs/current/indexam.html)
-- [Generic WAL Records](https://www.postgresql.org/docs/13/generic-wal.html)
+- [Generic WAL Records](https://www.postgresql.org/docs/current/generic-wal.html)

expected/cast.out

@@ -1,30 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-SELECT ARRAY[1,2,3]::vector;
-  array  
----------
- [1,2,3]
-(1 row)
-
-SELECT ARRAY[1,2,3]::float4[]::vector;
-  array  
----------
- [1,2,3]
-(1 row)
-
-SELECT ARRAY[1,2,3]::float8[]::vector;
-  array  
----------
- [1,2,3]
-(1 row)
-
-SELECT '{NULL}'::real[]::vector;
-ERROR:  array must not containing NULLs
-SELECT '{NaN}'::real[]::vector;
-ERROR:  NaN not allowed in vector
-SELECT '{Infinity}'::real[]::vector;
-ERROR:  infinite value not allowed in vector
-SELECT '{-Infinity}'::real[]::vector;
-ERROR:  infinite value not allowed in vector
-SELECT '{}'::real[]::vector;
-ERROR:  vector must have at least 1 dimension

expected/functions.out

@@ -1,56 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-SELECT '[1,2,3]'::vector + '[4,5,6]';
- ?column? 
-----------
- [5,7,9]
-(1 row)
-
-SELECT '[1,2,3]'::vector - '[4,5,6]';
-  ?column?  
-------------
- [-3,-3,-3]
-(1 row)
-
-SELECT vector_dims('[1,2,3]');
- vector_dims 
--------------
-           3
-(1 row)
-
-SELECT round(vector_norm('[1,1]')::numeric, 5);
-  round  
----------
- 1.41421
-(1 row)
-
-SELECT round(l2_distance('[1,2]', '[0,0]')::numeric, 5);
-  round  
----------
- 2.23607
-(1 row)
-
-SELECT l2_distance('[1,2]', '[3]');
-ERROR:  different vector dimensions 2 and 1
-SELECT inner_product('[1,2]', '[3,4]');
- inner_product 
----------------
-            11
-(1 row)
-
-SELECT inner_product('[1,2]', '[3]');
-ERROR:  different vector dimensions 2 and 1
-SELECT round(cosine_distance('[1,2]', '[2,4]')::numeric, 5);
-  round  
----------
- 0.00000
-(1 row)
-
-SELECT cosine_distance('[1,2]', '[0,0]');
- cosine_distance 
------------------
-             NaN
-(1 row)
-
-SELECT cosine_distance('[1,2]', '[3]');
-ERROR:  different vector dimensions 2 and 1

expected/vector.out

@@ -1,55 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-SELECT '[1,2,3]'::vector;
- vector  
----------
- [1,2,3]
-(1 row)
-
-SELECT '[-1,2,3]'::vector;
-  vector  
-----------
- [-1,2,3]
-(1 row)
-
-SELECT '[hello,1]'::vector;
-ERROR:  invalid input syntax for type vector: "hello"
-LINE 1: SELECT '[hello,1]'::vector;
-               ^
-SELECT '[NaN,1]'::vector;
-ERROR:  NaN not allowed in vector
-LINE 1: SELECT '[NaN,1]'::vector;
-               ^
-SELECT '[Infinity,1]'::vector;
-ERROR:  infinite value not allowed in vector
-LINE 1: SELECT '[Infinity,1]'::vector;
-               ^
-SELECT '[-Infinity,1]'::vector;
-ERROR:  infinite value not allowed in vector
-LINE 1: SELECT '[-Infinity,1]'::vector;
-               ^
-SELECT '[1,2,3'::vector;
-ERROR:  malformed vector literal
-LINE 1: SELECT '[1,2,3'::vector;
-               ^
-DETAIL:  Unexpected end of input.
-SELECT '[1,2,3]9'::vector;
-ERROR:  malformed vector literal
-LINE 1: SELECT '[1,2,3]9'::vector;
-               ^
-DETAIL:  Junk after closing right brace.
-SELECT '1,2,3'::vector;
-ERROR:  malformed vector literal: "1,2,3"
-LINE 1: SELECT '1,2,3'::vector;
-               ^
-DETAIL:  Vector contents must start with "[".
-SELECT '[]'::vector;
-ERROR:  vector must have at least 1 dimension
-LINE 1: SELECT '[]'::vector;
-               ^
-SELECT '[1,]'::vector;
-ERROR:  invalid input syntax for type vector: "]"
-LINE 1: SELECT '[1,]'::vector;
-               ^
-SELECT '[1,2,3]'::vector(2);
-ERROR:  expected 2 dimensions, not 3

sql/cast.sql

@@ -1,11 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-
-SELECT ARRAY[1,2,3]::vector;
-SELECT ARRAY[1,2,3]::float4[]::vector;
-SELECT ARRAY[1,2,3]::float8[]::vector;
-SELECT '{NULL}'::real[]::vector;
-SELECT '{NaN}'::real[]::vector;
-SELECT '{Infinity}'::real[]::vector;
-SELECT '{-Infinity}'::real[]::vector;
-SELECT '{}'::real[]::vector;

sql/functions.sql

@@ -1,18 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-
-SELECT '[1,2,3]'::vector + '[4,5,6]';
-SELECT '[1,2,3]'::vector - '[4,5,6]';
-
-SELECT vector_dims('[1,2,3]');
-SELECT round(vector_norm('[1,1]')::numeric, 5);
-
-SELECT round(l2_distance('[1,2]', '[0,0]')::numeric, 5);
-SELECT l2_distance('[1,2]', '[3]');
-
-SELECT inner_product('[1,2]', '[3,4]');
-SELECT inner_product('[1,2]', '[3]');
-
-SELECT round(cosine_distance('[1,2]', '[2,4]')::numeric, 5);
-SELECT cosine_distance('[1,2]', '[0,0]');
-SELECT cosine_distance('[1,2]', '[3]');

sql/ivfflat_options.sql

@@ -1,11 +0,0 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-SET enable_seqscan = off;
-
-CREATE TABLE t (val vector(3));
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
-
-SHOW ivfflat.probes;
-
-DROP TABLE t;

sql/vector--0.1.1--0.1.3.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.3'" to load this file. \quit

sql/vector--0.1.3--0.1.4.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.4'" to load this file. \quit

sql/vector--0.1.4--0.1.5.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.5'" to load this file. \quit

sql/vector--0.1.5--0.1.6.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.6'" to load this file. \quit

sql/vector--0.1.6--0.1.7.sql

@@ -0,0 +1,8 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.7'" to load this file. \quit
+
+CREATE FUNCTION array_to_vector(numeric[], integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE CAST (numeric[] AS vector)
+	WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS IMPLICIT;

sql/vector--0.1.7--0.1.8.sql

@@ -0,0 +1,8 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.1.8'" to load this file. \quit
+
+CREATE FUNCTION vector_to_float4(vector, integer, boolean) RETURNS real[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE CAST (vector AS real[])
+	WITH FUNCTION vector_to_float4(vector, integer, boolean) AS IMPLICIT;

sql/vector--0.1.8--0.2.0.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.0'" to load this file. \quit

sql/vector--0.2.0--0.2.1.sql

@@ -0,0 +1,19 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.1'" to load this file. \quit
+
+DROP CAST (integer[] AS vector);
+DROP CAST (real[] AS vector);
+DROP CAST (double precision[] AS vector);
+DROP CAST (numeric[] AS vector);
+
+CREATE CAST (integer[] AS vector)
+	WITH FUNCTION array_to_vector(integer[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (real[] AS vector)
+	WITH FUNCTION array_to_vector(real[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (double precision[] AS vector)
+	WITH FUNCTION array_to_vector(double precision[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (numeric[] AS vector)
+	WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;

sql/vector--0.2.1--0.2.2.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.2'" to load this file. \quit

sql/vector--0.2.2--0.2.3.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.3'" to load this file. \quit

sql/vector--0.2.3--0.2.4.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.4'" to load this file. \quit

sql/vector--0.2.4--0.2.5.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.5'" to load this file. \quit

sql/vector--0.2.5--0.2.6.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.6'" to load this file. \quit

sql/vector--0.2.6--0.2.7.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.2.7'" to load this file. \quit

sql/vector--0.2.7--0.3.0.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.3.0'" to load this file. \quit

sql/vector--0.3.0--0.3.1.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.3.1'" to load this file. \quit

sql/vector--0.3.1--0.3.2.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.3.2'" to load this file. \quit

sql/vector--0.3.2--0.4.0.sql

@@ -0,0 +1,23 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.4.0'" to load this file. \quit
+
+-- remove this single line for Postgres < 13
+ALTER TYPE vector SET (STORAGE = extended);
+
+CREATE FUNCTION vector_accum(double precision[], vector) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_avg(double precision[]) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_combine(double precision[], double precision[]) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE AGGREGATE avg(vector) (
+	SFUNC = vector_accum,
+	STYPE = double precision[],
+	FINALFUNC = vector_avg,
+	COMBINEFUNC = vector_combine,
+	INITCOND = '{0}',
+	PARALLEL = SAFE
+);

sql/vector--0.4.0--0.4.1.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.4.1'" to load this file. \quit

sql/vector--0.4.1--0.4.2.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.4.2'" to load this file. \quit

sql/vector--0.4.2--0.4.3.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.4.3'" to load this file. \quit

sql/vector--0.4.3--0.4.4.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.4.4'" to load this file. \quit

sql/vector--0.4.4--0.5.0.sql

@@ -0,0 +1,43 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.0'" to load this file. \quit
+
+CREATE FUNCTION l1_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE OPERATOR * (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_mul,
+	COMMUTATOR = *
+);
+
+CREATE AGGREGATE sum(vector) (
+	SFUNC = vector_add,
+	STYPE = vector,
+	COMBINEFUNC = vector_add,
+	PARALLEL = SAFE
+);
+
+CREATE FUNCTION hnswhandler(internal) RETURNS index_am_handler
+	AS 'MODULE_PATHNAME' LANGUAGE C;
+
+CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnswhandler;
+
+COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
+
+CREATE OPERATOR CLASS vector_l2_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <-> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_l2_squared_distance(vector, vector);
+
+CREATE OPERATOR CLASS vector_ip_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <#> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector);
+
+CREATE OPERATOR CLASS vector_cosine_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector),
+	FUNCTION 2 vector_norm(vector);

sql/vector.sql

@@ -1,15 +1,292 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "CREATE EXTENSION vector" to load this file. \quit
 
-SELECT '[1,2,3]'::vector;
-SELECT '[-1,2,3]'::vector;
-SELECT '[hello,1]'::vector;
-SELECT '[NaN,1]'::vector;
-SELECT '[Infinity,1]'::vector;
-SELECT '[-Infinity,1]'::vector;
-SELECT '[1,2,3'::vector;
-SELECT '[1,2,3]9'::vector;
-SELECT '1,2,3'::vector;
-SELECT '[]'::vector;
-SELECT '[1,]'::vector;
-SELECT '[1,2,3]'::vector(2);
+-- type
+
+CREATE TYPE vector;
+
+CREATE FUNCTION vector_in(cstring, oid, integer) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_out(vector) RETURNS cstring
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_typmod_in(cstring[]) RETURNS integer
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_recv(internal, oid, integer) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_send(vector) RETURNS bytea
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE TYPE vector (
+	INPUT     = vector_in,
+	OUTPUT    = vector_out,
+	TYPMOD_IN = vector_typmod_in,
+	RECEIVE   = vector_recv,
+	SEND      = vector_send,
+	STORAGE   = extended
+);
+
+-- functions
+
+CREATE FUNCTION l2_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION inner_product(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION cosine_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l1_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_dims(vector) RETURNS integer
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_norm(vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_add(vector, vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_sub(vector, vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- private functions
+
+CREATE FUNCTION vector_lt(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_le(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_eq(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_ne(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_ge(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_gt(vector, vector) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_cmp(vector, vector) RETURNS int4
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_l2_squared_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_negative_inner_product(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_spherical_distance(vector, vector) RETURNS float8
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_accum(double precision[], vector) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_avg(double precision[]) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_combine(double precision[], double precision[]) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- aggregates
+
+CREATE AGGREGATE avg(vector) (
+	SFUNC = vector_accum,
+	STYPE = double precision[],
+	FINALFUNC = vector_avg,
+	COMBINEFUNC = vector_combine,
+	INITCOND = '{0}',
+	PARALLEL = SAFE
+);
+
+CREATE AGGREGATE sum(vector) (
+	SFUNC = vector_add,
+	STYPE = vector,
+	COMBINEFUNC = vector_add,
+	PARALLEL = SAFE
+);
+
+-- cast functions
+
+CREATE FUNCTION vector(vector, integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION array_to_vector(integer[], integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION array_to_vector(real[], integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION array_to_vector(double precision[], integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION array_to_vector(numeric[], integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_to_float4(vector, integer, boolean) RETURNS real[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- casts
+
+CREATE CAST (vector AS vector)
+	WITH FUNCTION vector(vector, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (vector AS real[])
+	WITH FUNCTION vector_to_float4(vector, integer, boolean) AS IMPLICIT;
+
+CREATE CAST (integer[] AS vector)
+	WITH FUNCTION array_to_vector(integer[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (real[] AS vector)
+	WITH FUNCTION array_to_vector(real[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (double precision[] AS vector)
+	WITH FUNCTION array_to_vector(double precision[], integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (numeric[] AS vector)
+	WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;
+
+-- operators
+
+CREATE OPERATOR <-> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l2_distance,
+	COMMUTATOR = '<->'
+);
+
+CREATE OPERATOR <#> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_negative_inner_product,
+	COMMUTATOR = '<#>'
+);
+
+CREATE OPERATOR <=> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = cosine_distance,
+	COMMUTATOR = '<=>'
+);
+
+CREATE OPERATOR + (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_add,
+	COMMUTATOR = +
+);
+
+CREATE OPERATOR - (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub,
+	COMMUTATOR = -
+);
+
+CREATE OPERATOR * (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_mul,
+	COMMUTATOR = *
+);
+
+CREATE OPERATOR < (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_lt,
+	COMMUTATOR = > , NEGATOR = >= ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+-- should use scalarlesel and scalarlejoinsel, but not supported in Postgres < 11
+CREATE OPERATOR <= (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_le,
+	COMMUTATOR = >= , NEGATOR = > ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+CREATE OPERATOR = (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_eq,
+	COMMUTATOR = = , NEGATOR = <> ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR <> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_ne,
+	COMMUTATOR = <> , NEGATOR = = ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+-- should use scalargesel and scalargejoinsel, but not supported in Postgres < 11
+CREATE OPERATOR >= (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_ge,
+	COMMUTATOR = <= , NEGATOR = < ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
+CREATE OPERATOR > (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_gt,
+	COMMUTATOR = < , NEGATOR = <= ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
+-- access methods
+
+CREATE FUNCTION ivfflathandler(internal) RETURNS index_am_handler
+	AS 'MODULE_PATHNAME' LANGUAGE C;
+
+CREATE ACCESS METHOD ivfflat TYPE INDEX HANDLER ivfflathandler;
+
+COMMENT ON ACCESS METHOD ivfflat IS 'ivfflat index access method';
+
+CREATE FUNCTION hnswhandler(internal) RETURNS index_am_handler
+	AS 'MODULE_PATHNAME' LANGUAGE C;
+
+CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnswhandler;
+
+COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
+
+-- opclasses
+
+CREATE OPERATOR CLASS vector_ops
+	DEFAULT FOR TYPE vector USING btree AS
+	OPERATOR 1 < ,
+	OPERATOR 2 <= ,
+	OPERATOR 3 = ,
+	OPERATOR 4 >= ,
+	OPERATOR 5 > ,
+	FUNCTION 1 vector_cmp(vector, vector);
+
+CREATE OPERATOR CLASS vector_l2_ops
+	DEFAULT FOR TYPE vector USING ivfflat AS
+	OPERATOR 1 <-> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_l2_squared_distance(vector, vector),
+	FUNCTION 3 l2_distance(vector, vector);
+
+CREATE OPERATOR CLASS vector_ip_ops
+	FOR TYPE vector USING ivfflat AS
+	OPERATOR 1 <#> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector),
+	FUNCTION 3 vector_spherical_distance(vector, vector),
+	FUNCTION 4 vector_norm(vector);
+
+CREATE OPERATOR CLASS vector_cosine_ops
+	FOR TYPE vector USING ivfflat AS
+	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector),
+	FUNCTION 2 vector_norm(vector),
+	FUNCTION 3 vector_spherical_distance(vector, vector),
+	FUNCTION 4 vector_norm(vector);
+
+CREATE OPERATOR CLASS vector_l2_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <-> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_l2_squared_distance(vector, vector);
+
+CREATE OPERATOR CLASS vector_ip_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <#> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector);
+
+CREATE OPERATOR CLASS vector_cosine_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 vector_negative_inner_product(vector, vector),
+	FUNCTION 2 vector_norm(vector);

src/hnsw.c

@@ -0,0 +1,224 @@
+#include "postgres.h"
+
+#include <float.h>
+#include <math.h>
+
+#include "access/amapi.h"
+#include "commands/vacuum.h"
+#include "hnsw.h"
+#include "utils/guc.h"
+#include "utils/selfuncs.h"
+
+#if PG_VERSION_NUM >= 120000
+#include "commands/progress.h"
+#endif
+
+int			hnsw_ef_search;
+static relopt_kind hnsw_relopt_kind;
+
+/*
+ * Initialize index options and variables
+ */
+void
+HnswInit(void)
+{
+	hnsw_relopt_kind = add_reloption_kind();
+	add_int_reloption(hnsw_relopt_kind, "m", "Max number of connections",
+					  HNSW_DEFAULT_M, HNSW_MIN_M, HNSW_MAX_M
+#if PG_VERSION_NUM >= 130000
+					  ,AccessExclusiveLock
+#endif
+		);
+	add_int_reloption(hnsw_relopt_kind, "ef_construction", "Size of the dynamic candidate list for construction",
+					  HNSW_DEFAULT_EF_CONSTRUCTION, HNSW_MIN_EF_CONSTRUCTION, HNSW_MAX_EF_CONSTRUCTION
+#if PG_VERSION_NUM >= 130000
+					  ,AccessExclusiveLock
+#endif
+		);
+
+	DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
+							"Valid range is 1..1000.", &hnsw_ef_search,
+							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
+}
+
+/*
+ * Get the name of index build phase
+ */
+#if PG_VERSION_NUM >= 120000
+static char *
+hnswbuildphasename(int64 phasenum)
+{
+	switch (phasenum)
+	{
+		case PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE:
+			return "initializing";
+		case PROGRESS_HNSW_PHASE_LOAD:
+			return "loading tuples";
+		default:
+			return NULL;
+	}
+}
+#endif
+
+/*
+ * Estimate the cost of an index scan
+ */
+static void
+hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
+				 Cost *indexStartupCost, Cost *indexTotalCost,
+				 Selectivity *indexSelectivity, double *indexCorrelation,
+				 double *indexPages)
+{
+	GenericCosts costs;
+	int			m;
+	int			entryLevel;
+	Relation	index;
+#if PG_VERSION_NUM < 120000
+	List	   *qinfos;
+#endif
+
+	/* Never use index without order */
+	if (path->indexorderbys == NULL)
+	{
+		*indexStartupCost = DBL_MAX;
+		*indexTotalCost = DBL_MAX;
+		*indexSelectivity = 0;
+		*indexCorrelation = 0;
+		*indexPages = 0;
+		return;
+	}
+
+	MemSet(&costs, 0, sizeof(costs));
+
+	index = index_open(path->indexinfo->indexoid, NoLock);
+	HnswGetMetaPageInfo(index, &m, NULL);
+	index_close(index, NoLock);
+
+	/* Approximate entry level */
+	entryLevel = (int) -log(1.0 / path->indexinfo->tuples) * HnswGetMl(m);
+
+	/* TODO Improve estimate of visited tuples (currently underestimates) */
+	/* Account for number of tuples (or entry level), m, and ef_search */
+	costs.numIndexTuples = (entryLevel + 2) * m;
+
+#if PG_VERSION_NUM >= 120000
+	genericcostestimate(root, path, loop_count, &costs);
+#else
+	qinfos = deconstruct_indexquals(path);
+	genericcostestimate(root, path, loop_count, qinfos, &costs);
+#endif
+
+	/* Use total cost since most work happens before first tuple is returned */
+	*indexStartupCost = costs.indexTotalCost;
+	*indexTotalCost = costs.indexTotalCost;
+	*indexSelectivity = costs.indexSelectivity;
+	*indexCorrelation = costs.indexCorrelation;
+	*indexPages = costs.numIndexPages;
+}
+
+/*
+ * Parse and validate the reloptions
+ */
+static bytea *
+hnswoptions(Datum reloptions, bool validate)
+{
+	static const relopt_parse_elt tab[] = {
+		{"m", RELOPT_TYPE_INT, offsetof(HnswOptions, m)},
+		{"ef_construction", RELOPT_TYPE_INT, offsetof(HnswOptions, efConstruction)},
+	};
+
+#if PG_VERSION_NUM >= 130000
+	return (bytea *) build_reloptions(reloptions, validate,
+									  hnsw_relopt_kind,
+									  sizeof(HnswOptions),
+									  tab, lengthof(tab));
+#else
+	relopt_value *options;
+	int			numoptions;
+	HnswOptions *rdopts;
+
+	options = parseRelOptions(reloptions, validate, hnsw_relopt_kind, &numoptions);
+	rdopts = allocateReloptStruct(sizeof(HnswOptions), options, numoptions);
+	fillRelOptions((void *) rdopts, sizeof(HnswOptions), options, numoptions,
+				   validate, tab, lengthof(tab));
+
+	return (bytea *) rdopts;
+#endif
+}
+
+/*
+ * Validate catalog entries for the specified operator class
+ */
+static bool
+hnswvalidate(Oid opclassoid)
+{
+	return true;
+}
+
+/*
+ * Define index handler
+ *
+ * See https://www.postgresql.org/docs/current/index-api.html
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(hnswhandler);
+Datum
+hnswhandler(PG_FUNCTION_ARGS)
+{
+	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
+
+	amroutine->amstrategies = 0;
+	amroutine->amsupport = 2;
+#if PG_VERSION_NUM >= 130000
+	amroutine->amoptsprocnum = 0;
+#endif
+	amroutine->amcanorder = false;
+	amroutine->amcanorderbyop = true;
+	amroutine->amcanbackward = false;	/* can change direction mid-scan */
+	amroutine->amcanunique = false;
+	amroutine->amcanmulticol = false;
+	amroutine->amoptionalkey = true;
+	amroutine->amsearcharray = false;
+	amroutine->amsearchnulls = false;
+	amroutine->amstorage = false;
+	amroutine->amclusterable = false;
+	amroutine->ampredlocks = false;
+	amroutine->amcanparallel = false;
+	amroutine->amcaninclude = false;
+#if PG_VERSION_NUM >= 130000
+	amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
+	amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
+#endif
+	amroutine->amkeytype = InvalidOid;
+
+	/* Interface functions */
+	amroutine->ambuild = hnswbuild;
+	amroutine->ambuildempty = hnswbuildempty;
+	amroutine->aminsert = hnswinsert;
+	amroutine->ambulkdelete = hnswbulkdelete;
+	amroutine->amvacuumcleanup = hnswvacuumcleanup;
+	amroutine->amcanreturn = NULL;
+	amroutine->amcostestimate = hnswcostestimate;
+	amroutine->amoptions = hnswoptions;
+	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */
+#if PG_VERSION_NUM >= 120000
+	amroutine->ambuildphasename = hnswbuildphasename;
+#endif
+	amroutine->amvalidate = hnswvalidate;
+#if PG_VERSION_NUM >= 140000
+	amroutine->amadjustmembers = NULL;
+#endif
+	amroutine->ambeginscan = hnswbeginscan;
+	amroutine->amrescan = hnswrescan;
+	amroutine->amgettuple = hnswgettuple;
+	amroutine->amgetbitmap = NULL;
+	amroutine->amendscan = hnswendscan;
+	amroutine->ammarkpos = NULL;
+	amroutine->amrestrpos = NULL;
+
+	/* Interface functions to support parallel index scans */
+	amroutine->amestimateparallelscan = NULL;
+	amroutine->aminitparallelscan = NULL;
+	amroutine->amparallelrescan = NULL;
+
+	PG_RETURN_POINTER(amroutine);
+}

src/hnsw.h

@@ -0,0 +1,305 @@
+#ifndef HNSW_H
+#define HNSW_H
+
+#include "postgres.h"
+
+#include "access/generic_xlog.h"
+#include "access/reloptions.h"
+#include "nodes/execnodes.h"
+#include "port.h"				/* for random() */
+#include "utils/sampling.h"
+#include "vector.h"
+
+#if PG_VERSION_NUM < 110000
+#error "Requires PostgreSQL 11+"
+#endif
+
+#define HNSW_MAX_DIM 2000
+
+/* Support functions */
+#define HNSW_DISTANCE_PROC 1
+#define HNSW_NORM_PROC 2
+
+#define HNSW_VERSION	1
+#define HNSW_MAGIC_NUMBER 0xA953A953
+#define HNSW_PAGE_ID	0xFF90
+
+/* Preserved page numbers */
+#define HNSW_METAPAGE_BLKNO	0
+#define HNSW_HEAD_BLKNO		1	/* first element page */
+
+/* Must correspond to page numbers since page lock is used */
+#define HNSW_UPDATE_LOCK 	0
+#define HNSW_SCAN_LOCK		1
+
+/* HNSW parameters */
+#define HNSW_DEFAULT_M	16
+#define HNSW_MIN_M	2
+#define HNSW_MAX_M		100
+#define HNSW_DEFAULT_EF_CONSTRUCTION	64
+#define HNSW_MIN_EF_CONSTRUCTION	4
+#define HNSW_MAX_EF_CONSTRUCTION		1000
+#define HNSW_DEFAULT_EF_SEARCH	40
+#define HNSW_MIN_EF_SEARCH		1
+#define HNSW_MAX_EF_SEARCH		1000
+
+/* Tuple types */
+#define HNSW_ELEMENT_TUPLE_TYPE  1
+#define HNSW_NEIGHBOR_TUPLE_TYPE 2
+
+/* Make graph robust against non-HOT updates */
+#define HNSW_HEAPTIDS 10
+
+#define HNSW_UPDATE_ENTRY_GREATER 1
+#define HNSW_UPDATE_ENTRY_ALWAYS 2
+
+/* Build phases */
+/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
+#define PROGRESS_HNSW_PHASE_LOAD		2
+
+#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
+#define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
+
+#define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
+#define HnswPageGetMeta(page)	((HnswMetaPageData *) PageGetContents(page))
+
+#if PG_VERSION_NUM >= 150000
+#define RandomDouble() pg_prng_double(&pg_global_prng_state)
+#else
+#define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
+#endif
+
+#if PG_VERSION_NUM < 130000
+#define list_delete_last(list) list_truncate(list, list_length(list) - 1)
+#define list_sort(list, cmp) list_qsort(list, cmp)
+#endif
+
+#define HnswIsElementTuple(tup) ((tup)->type == HNSW_ELEMENT_TUPLE_TYPE)
+#define HnswIsNeighborTuple(tup) ((tup)->type == HNSW_NEIGHBOR_TUPLE_TYPE)
+
+/* 2 * M connections for ground layer */
+#define HnswGetLayerM(m, layer) (layer == 0 ? (m) * 2 : (m))
+
+/* Optimal ML from paper */
+#define HnswGetMl(m) (1 / log(m))
+
+/* Ensure fits on page and in uint8 */
+#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / m) - 2, 255)
+
+/* Variables */
+extern int	hnsw_ef_search;
+
+typedef struct HnswNeighborArray HnswNeighborArray;
+
+typedef struct HnswElementData
+{
+	List	   *heaptids;
+	uint8		level;
+	uint8		deleted;
+	HnswNeighborArray *neighbors;
+	BlockNumber blkno;
+	OffsetNumber offno;
+	OffsetNumber neighborOffno;
+	BlockNumber neighborPage;
+	Vector	   *vec;
+}			HnswElementData;
+
+typedef HnswElementData * HnswElement;
+
+typedef struct HnswCandidate
+{
+	HnswElement element;
+	float		distance;
+}			HnswCandidate;
+
+typedef struct HnswNeighborArray
+{
+	int			length;
+	HnswCandidate *items;
+}			HnswNeighborArray;
+
+typedef struct HnswPairingHeapNode
+{
+	pairingheap_node ph_node;
+	HnswCandidate *inner;
+}			HnswPairingHeapNode;
+
+/* HNSW index options */
+typedef struct HnswOptions
+{
+	int32		vl_len_;		/* varlena header (do not touch directly!) */
+	int			m;				/* number of connections */
+	int			efConstruction; /* size of dynamic candidate list */
+}			HnswOptions;
+
+typedef struct HnswBuildState
+{
+	/* Info */
+	Relation	heap;
+	Relation	index;
+	IndexInfo  *indexInfo;
+	ForkNumber	forkNum;
+
+	/* Settings */
+	int			dimensions;
+	int			m;
+	int			efConstruction;
+
+	/* Statistics */
+	double		indtuples;
+	double		reltuples;
+
+	/* Support functions */
+	FmgrInfo   *procinfo;
+	FmgrInfo   *normprocinfo;
+	Oid			collation;
+
+	/* Variables */
+	List	   *elements;
+	HnswElement entryPoint;
+	double		ml;
+	int			maxLevel;
+	double		maxInMemoryElements;
+	bool		flushed;
+	Vector	   *normvec;
+
+	/* Memory */
+	MemoryContext tmpCtx;
+}			HnswBuildState;
+
+typedef struct HnswMetaPageData
+{
+	uint32		magicNumber;
+	uint32		version;
+	uint32		dimensions;
+	uint16		m;
+	uint16		efConstruction;
+	BlockNumber entryBlkno;
+	OffsetNumber entryOffno;
+	int16		entryLevel;
+	BlockNumber insertPage;
+}			HnswMetaPageData;
+
+typedef HnswMetaPageData * HnswMetaPage;
+
+typedef struct HnswPageOpaqueData
+{
+	BlockNumber nextblkno;
+	uint16		unused;
+	uint16		page_id;		/* for identification of HNSW indexes */
+}			HnswPageOpaqueData;
+
+typedef HnswPageOpaqueData * HnswPageOpaque;
+
+typedef struct HnswElementTupleData
+{
+	uint8		type;
+	uint8		level;
+	uint8		deleted;
+	uint8		unused;
+	ItemPointerData heaptids[HNSW_HEAPTIDS];
+	ItemPointerData neighbortid;
+	uint16		unused2;
+	Vector		vec;
+}			HnswElementTupleData;
+
+typedef HnswElementTupleData * HnswElementTuple;
+
+typedef struct HnswNeighborTupleData
+{
+	uint8		type;
+	uint8		unused;
+	uint16		count;
+	ItemPointerData indextids[FLEXIBLE_ARRAY_MEMBER];
+}			HnswNeighborTupleData;
+
+typedef HnswNeighborTupleData * HnswNeighborTuple;
+
+typedef struct HnswScanOpaqueData
+{
+	bool		first;
+	List	   *w;
+	MemoryContext tmpCtx;
+
+	/* Support functions */
+	FmgrInfo   *procinfo;
+	FmgrInfo   *normprocinfo;
+	Oid			collation;
+}			HnswScanOpaqueData;
+
+typedef HnswScanOpaqueData * HnswScanOpaque;
+
+typedef struct HnswVacuumState
+{
+	/* Info */
+	Relation	index;
+	IndexBulkDeleteResult *stats;
+	IndexBulkDeleteCallback callback;
+	void	   *callback_state;
+
+	/* Settings */
+	int			m;
+	int			efConstruction;
+
+	/* Support functions */
+	FmgrInfo   *procinfo;
+	Oid			collation;
+
+	/* Variables */
+	HTAB	   *deleted;
+	BufferAccessStrategy bas;
+	HnswNeighborTuple ntup;
+	HnswElementData highestPoint;
+
+	/* Memory */
+	MemoryContext tmpCtx;
+}			HnswVacuumState;
+
+/* Methods */
+int			HnswGetM(Relation index);
+int			HnswGetEfConstruction(Relation index);
+FmgrInfo   *HnswOptionalProcInfo(Relation rel, uint16 procnum);
+bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+void		HnswCommitBuffer(Buffer buf, GenericXLogState *state);
+Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
+void		HnswInitPage(Buffer buf, Page page);
+void		HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
+void		HnswInit(void);
+List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
+HnswElement HnswGetEntryPoint(Relation index);
+void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
+HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
+void		HnswFreeElement(HnswElement element);
+HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
+void		HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
+HnswElement HnswFindDuplicate(HnswElement e);
+HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
+void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum);
+void		HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m);
+void		HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
+void		HnswInitNeighbors(HnswElement element, int m);
+bool		HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel);
+void		HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting);
+void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
+void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec);
+void		HnswSetElementTuple(HnswElementTuple etup, HnswElement element);
+void		HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
+void		HnswLoadNeighbors(HnswElement element, Relation index, int m);
+
+/* Index access methods */
+IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);
+void		hnswbuildempty(Relation index);
+bool		hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heap, IndexUniqueCheck checkUnique
+#if PG_VERSION_NUM >= 140000
+					   ,bool indexUnchanged
+#endif
+					   ,IndexInfo *indexInfo
+);
+IndexBulkDeleteResult *hnswbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats, IndexBulkDeleteCallback callback, void *callback_state);
+IndexBulkDeleteResult *hnswvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats);
+IndexScanDesc hnswbeginscan(Relation index, int nkeys, int norderbys);
+void		hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int norderbys);
+bool		hnswgettuple(IndexScanDesc scan, ScanDirection dir);
+void		hnswendscan(IndexScanDesc scan);
+
+#endif

src/hnswbuild.c

@@ -0,0 +1,526 @@
+#include "postgres.h"
+
+#include <math.h>
+
+#include "catalog/index.h"
+#include "hnsw.h"
+#include "miscadmin.h"
+#include "lib/pairingheap.h"
+#include "nodes/pg_list.h"
+#include "storage/bufmgr.h"
+#include "utils/memutils.h"
+
+#if PG_VERSION_NUM >= 140000
+#include "utils/backend_progress.h"
+#elif PG_VERSION_NUM >= 120000
+#include "pgstat.h"
+#endif
+
+#if PG_VERSION_NUM >= 120000
+#include "access/tableam.h"
+#include "commands/progress.h"
+#else
+#define PROGRESS_CREATEIDX_TUPLES_DONE 0
+#endif
+
+#if PG_VERSION_NUM >= 130000
+#define CALLBACK_ITEM_POINTER ItemPointer tid
+#else
+#define CALLBACK_ITEM_POINTER HeapTuple hup
+#endif
+
+#if PG_VERSION_NUM >= 120000
+#define UpdateProgress(index, val) pgstat_progress_update_param(index, val)
+#else
+#define UpdateProgress(index, val) ((void)val)
+#endif
+
+/*
+ * Create the metapage
+ */
+static void
+CreateMetaPage(HnswBuildState * buildstate)
+{
+	Relation	index = buildstate->index;
+	ForkNumber	forkNum = buildstate->forkNum;
+	Buffer		buf;
+	Page		page;
+	GenericXLogState *state;
+	HnswMetaPage metap;
+
+	buf = HnswNewBuffer(index, forkNum);
+	HnswInitRegisterPage(index, &buf, &page, &state);
+
+	/* Set metapage data */
+	metap = HnswPageGetMeta(page);
+	metap->magicNumber = HNSW_MAGIC_NUMBER;
+	metap->version = HNSW_VERSION;
+	metap->dimensions = buildstate->dimensions;
+	metap->m = buildstate->m;
+	metap->efConstruction = buildstate->efConstruction;
+	metap->entryBlkno = InvalidBlockNumber;
+	metap->entryOffno = InvalidOffsetNumber;
+	metap->entryLevel = -1;
+	metap->insertPage = InvalidBlockNumber;
+	((PageHeader) page)->pd_lower =
+		((char *) metap + sizeof(HnswMetaPageData)) - (char *) page;
+
+	HnswCommitBuffer(buf, state);
+}
+
+/*
+ * Add a new page
+ */
+static void
+HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum)
+{
+	/* Add a new page */
+	Buffer		newbuf = HnswNewBuffer(index, forkNum);
+
+	/* Update previous page */
+	HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
+
+	/* Commit */
+	MarkBufferDirty(*buf);
+	GenericXLogFinish(*state);
+	UnlockReleaseBuffer(*buf);
+
+	/* Can take a while, so ensure we can interrupt */
+	/* Needs to be called when no buffer locks are held */
+	LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
+	CHECK_FOR_INTERRUPTS();
+	LockBuffer(newbuf, BUFFER_LOCK_EXCLUSIVE);
+
+	/* Prepare new page */
+	*buf = newbuf;
+	*state = GenericXLogStart(index);
+	*page = GenericXLogRegisterBuffer(*state, *buf, GENERIC_XLOG_FULL_IMAGE);
+	HnswInitPage(*buf, *page);
+}
+
+/*
+ * Create element pages
+ */
+static void
+CreateElementPages(HnswBuildState * buildstate)
+{
+	Relation	index = buildstate->index;
+	ForkNumber	forkNum = buildstate->forkNum;
+	int			dimensions = buildstate->dimensions;
+	Size		etupSize;
+	Size		maxSize;
+	HnswElementTuple etup;
+	HnswNeighborTuple ntup;
+	BlockNumber insertPage;
+	Buffer		buf;
+	Page		page;
+	GenericXLogState *state;
+	ListCell   *lc;
+
+	/* Calculate sizes */
+	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+
+	/* Allocate once */
+	etup = palloc0(etupSize);
+	ntup = palloc0(maxSize);
+
+	/* Prepare first page */
+	buf = HnswNewBuffer(index, forkNum);
+	state = GenericXLogStart(index);
+	page = GenericXLogRegisterBuffer(state, buf, GENERIC_XLOG_FULL_IMAGE);
+	HnswInitPage(buf, page);
+
+	foreach(lc, buildstate->elements)
+	{
+		HnswElement element = lfirst(lc);
+		Size		ntupSize;
+		Size		combinedSize;
+
+		HnswSetElementTuple(etup, element);
+
+		/* Calculate sizes */
+		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
+		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
+
+		/* Keep element and neighbors on the same page if possible */
+		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
+			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
+
+		/* Calculate offsets */
+		element->blkno = BufferGetBlockNumber(buf);
+		element->offno = OffsetNumberNext(PageGetMaxOffsetNumber(page));
+		if (combinedSize <= maxSize)
+		{
+			element->neighborPage = element->blkno;
+			element->neighborOffno = OffsetNumberNext(element->offno);
+		}
+		else
+		{
+			element->neighborPage = element->blkno + 1;
+			element->neighborOffno = FirstOffsetNumber;
+		}
+
+		ItemPointerSet(&etup->neighbortid, element->neighborPage, element->neighborOffno);
+
+		/* Add element */
+		if (PageAddItem(page, (Item) etup, etupSize, InvalidOffsetNumber, false, false) != element->offno)
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+		/* Add new page if needed */
+		if (PageGetFreeSpace(page) < ntupSize)
+			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
+
+		/* Add placeholder for neighbors */
+		if (PageAddItem(page, (Item) ntup, ntupSize, InvalidOffsetNumber, false, false) != element->neighborOffno)
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+	}
+
+	insertPage = BufferGetBlockNumber(buf);
+
+	/* Commit */
+	MarkBufferDirty(buf);
+	GenericXLogFinish(state);
+	UnlockReleaseBuffer(buf);
+
+	HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, buildstate->entryPoint, insertPage, forkNum);
+
+	pfree(etup);
+	pfree(ntup);
+}
+
+/*
+ * Create neighbor pages
+ */
+static void
+CreateNeighborPages(HnswBuildState * buildstate)
+{
+	Relation	index = buildstate->index;
+	ForkNumber	forkNum = buildstate->forkNum;
+	int			m = buildstate->m;
+	ListCell   *lc;
+	HnswNeighborTuple ntup;
+
+	/* Allocate once */
+	ntup = palloc0(BLCKSZ);
+
+	foreach(lc, buildstate->elements)
+	{
+		HnswElement e = lfirst(lc);
+		Buffer		buf;
+		Page		page;
+		GenericXLogState *state;
+		Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
+
+		/* Can take a while, so ensure we can interrupt */
+		/* Needs to be called when no buffer locks are held */
+		CHECK_FOR_INTERRUPTS();
+
+		buf = ReadBufferExtended(index, forkNum, e->neighborPage, RBM_NORMAL, NULL);
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+
+		HnswSetNeighborTuple(ntup, e, m);
+
+		if (!PageIndexTupleOverwrite(page, e->neighborOffno, (Item) ntup, ntupSize))
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+		/* Commit */
+		MarkBufferDirty(buf);
+		GenericXLogFinish(state);
+		UnlockReleaseBuffer(buf);
+	}
+
+	pfree(ntup);
+}
+
+/*
+ * Free elements
+ */
+static void
+FreeElements(HnswBuildState * buildstate)
+{
+	ListCell   *lc;
+
+	foreach(lc, buildstate->elements)
+		HnswFreeElement(lfirst(lc));
+
+	list_free(buildstate->elements);
+}
+
+/*
+ * Flush pages
+ */
+static void
+FlushPages(HnswBuildState * buildstate)
+{
+	CreateMetaPage(buildstate);
+	CreateElementPages(buildstate);
+	CreateNeighborPages(buildstate);
+
+	buildstate->flushed = true;
+	FreeElements(buildstate);
+}
+
+/*
+ * Insert tuple
+ */
+static bool
+InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup)
+{
+	FmgrInfo   *procinfo = buildstate->procinfo;
+	Oid			collation = buildstate->collation;
+	HnswElement entryPoint = buildstate->entryPoint;
+	int			efConstruction = buildstate->efConstruction;
+	int			m = buildstate->m;
+
+	/* Detoast once for all calls */
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Normalize if needed */
+	if (buildstate->normprocinfo != NULL)
+	{
+		if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
+			return false;
+	}
+
+	/* Copy value to element so accessible outside of memory context */
+	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
+
+	/* Insert element in graph */
+	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
+
+	/* Look for duplicate */
+	*dup = HnswFindDuplicate(element);
+
+	/* Update neighbors if needed */
+	if (*dup == NULL)
+	{
+		for (int lc = element->level; lc >= 0; lc--)
+		{
+			int			lm = HnswGetLayerM(m, lc);
+			HnswNeighborArray *neighbors = &element->neighbors[lc];
+
+			for (int i = 0; i < neighbors->length; i++)
+				HnswUpdateConnection(element, &neighbors->items[i], lm, lc, NULL, NULL, procinfo, collation);
+		}
+	}
+
+	/* Update entry point if needed */
+	if (*dup == NULL && (entryPoint == NULL || element->level > entryPoint->level))
+		buildstate->entryPoint = element;
+
+	UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+
+	return *dup == NULL;
+}
+
+/*
+ * Callback for table_index_build_scan
+ */
+static void
+BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
+			  bool *isnull, bool tupleIsAlive, void *state)
+{
+	HnswBuildState *buildstate = (HnswBuildState *) state;
+	MemoryContext oldCtx;
+	HnswElement element;
+	HnswElement dup = NULL;
+	bool		inserted;
+
+#if PG_VERSION_NUM < 130000
+	ItemPointer tid = &hup->t_self;
+#endif
+
+	/* Skip nulls */
+	if (isnull[0])
+		return;
+
+	if (buildstate->indtuples >= buildstate->maxInMemoryElements)
+	{
+		if (!buildstate->flushed)
+		{
+			ereport(NOTICE,
+					(errmsg("hnsw graph no longer fits into maintenance_work_mem after " INT64_FORMAT " tuples", (int64) buildstate->indtuples),
+					 errdetail("Building will take significantly more time."),
+					 errhint("Increase maintenance_work_mem to speed up builds.")));
+
+			FlushPages(buildstate);
+		}
+
+		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
+
+		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
+			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+
+		/* Reset memory context */
+		MemoryContextSwitchTo(oldCtx);
+		MemoryContextReset(buildstate->tmpCtx);
+
+		return;
+	}
+
+	/* Allocate necessary memory outside of memory context */
+	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
+	element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
+
+	/* Use memory context since detoast can allocate */
+	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
+
+	/* Insert tuple */
+	inserted = InsertTuple(index, values, element, buildstate, &dup);
+
+	/* Reset memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextReset(buildstate->tmpCtx);
+
+	/* Add outside memory context */
+	if (dup != NULL)
+		HnswAddHeapTid(dup, tid);
+
+	/* Add to buildstate or free */
+	if (inserted)
+		buildstate->elements = lappend(buildstate->elements, element);
+	else
+		HnswFreeElement(element);
+}
+
+/*
+ * Get the max number of elements that fit into maintenance_work_mem
+ */
+static double
+HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
+{
+	Size		elementSize = sizeof(HnswElementData);
+	double		avgLevel = -log(0.5) * ml;
+
+	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
+	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
+	elementSize += sizeof(ItemPointerData);
+	elementSize += VECTOR_SIZE(dimensions);
+	return (maintenance_work_mem * 1024L) / elementSize;
+}
+
+/*
+ * Initialize the build state
+ */
+static void
+InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo, ForkNumber forkNum)
+{
+	buildstate->heap = heap;
+	buildstate->index = index;
+	buildstate->indexInfo = indexInfo;
+	buildstate->forkNum = forkNum;
+
+	buildstate->m = HnswGetM(index);
+	buildstate->efConstruction = HnswGetEfConstruction(index);
+	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
+
+	/* Require column to have dimensions to be indexed */
+	if (buildstate->dimensions < 0)
+		elog(ERROR, "column does not have dimensions");
+
+	if (buildstate->dimensions > HNSW_MAX_DIM)
+		elog(ERROR, "column cannot have more than %d dimensions for hnsw index", HNSW_MAX_DIM);
+
+	if (buildstate->efConstruction < 2 * buildstate->m)
+		elog(ERROR, "ef_construction must be greater than or equal to 2 * m");
+
+	buildstate->reltuples = 0;
+	buildstate->indtuples = 0;
+
+	/* Get support functions */
+	buildstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	buildstate->collation = index->rd_indcollation[0];
+
+	buildstate->elements = NIL;
+	buildstate->entryPoint = NULL;
+	buildstate->ml = HnswGetMl(buildstate->m);
+	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
+	buildstate->maxInMemoryElements = HnswGetMaxInMemoryElements(buildstate->m, buildstate->ml, buildstate->dimensions);
+	buildstate->flushed = false;
+
+	/* Reuse for each tuple */
+	buildstate->normvec = InitVector(buildstate->dimensions);
+
+	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+											   "Hnsw build temporary context",
+											   ALLOCSET_DEFAULT_SIZES);
+}
+
+/*
+ * Free resources
+ */
+static void
+FreeBuildState(HnswBuildState * buildstate)
+{
+	pfree(buildstate->normvec);
+	MemoryContextDelete(buildstate->tmpCtx);
+}
+
+/*
+ * Build graph
+ */
+static void
+BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
+{
+	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_HNSW_PHASE_LOAD);
+
+#if PG_VERSION_NUM >= 120000
+	buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+												   true, true, BuildCallback, (void *) buildstate, NULL);
+#else
+	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+											   true, BuildCallback, (void *) buildstate, NULL);
+#endif
+}
+
+/*
+ * Build the index
+ */
+static void
+BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
+		   HnswBuildState * buildstate, ForkNumber forkNum)
+{
+	InitBuildState(buildstate, heap, index, indexInfo, forkNum);
+
+	if (buildstate->heap != NULL)
+		BuildGraph(buildstate, forkNum);
+
+	if (!buildstate->flushed)
+		FlushPages(buildstate);
+
+	FreeBuildState(buildstate);
+}
+
+/*
+ * Build the index for a logged table
+ */
+IndexBuildResult *
+hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo)
+{
+	IndexBuildResult *result;
+	HnswBuildState buildstate;
+
+	BuildIndex(heap, index, indexInfo, &buildstate, MAIN_FORKNUM);
+
+	result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
+	result->heap_tuples = buildstate.reltuples;
+	result->index_tuples = buildstate.indtuples;
+
+	return result;
+}
+
+/*
+ * Build the index for an unlogged table
+ */
+void
+hnswbuildempty(Relation index)
+{
+	IndexInfo  *indexInfo = BuildIndexInfo(index);
+	HnswBuildState buildstate;
+
+	BuildIndex(NULL, index, indexInfo, &buildstate, INIT_FORKNUM);
+}

src/hnswinsert.c

@@ -0,0 +1,589 @@
+#include "postgres.h"
+
+#include <math.h>
+
+#include "hnsw.h"
+#include "storage/bufmgr.h"
+#include "storage/lmgr.h"
+#include "utils/memutils.h"
+
+/*
+ * Get the insert page
+ */
+static BlockNumber
+GetInsertPage(Relation index)
+{
+	Buffer		buf;
+	Page		page;
+	HnswMetaPage metap;
+	BlockNumber insertPage;
+
+	buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	metap = HnswPageGetMeta(page);
+
+	insertPage = metap->insertPage;
+
+	UnlockReleaseBuffer(buf);
+
+	return insertPage;
+}
+
+/*
+ * Check for a free offset
+ */
+static bool
+HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage)
+{
+	OffsetNumber offno;
+	OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
+
+	for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+	{
+		HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+
+		/* Skip neighbor tuples */
+		if (!HnswIsElementTuple(etup))
+			continue;
+
+		if (etup->deleted)
+		{
+			BlockNumber elementPage = BufferGetBlockNumber(buf);
+			BlockNumber neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
+			OffsetNumber neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
+			ItemId		itemid;
+
+			if (!BlockNumberIsValid(*newInsertPage))
+				*newInsertPage = elementPage;
+
+			if (neighborPage == elementPage)
+			{
+				*nbuf = buf;
+				*npage = page;
+			}
+			else
+			{
+				*nbuf = ReadBuffer(index, neighborPage);
+				LockBuffer(*nbuf, BUFFER_LOCK_EXCLUSIVE);
+
+				/* Skip WAL for now */
+				*npage = BufferGetPage(*nbuf);
+			}
+
+			itemid = PageGetItemId(*npage, neighborOffno);
+
+			/* Check for space on neighbor tuple page */
+			if (PageGetFreeSpace(*npage) + ItemIdGetLength(itemid) - sizeof(ItemIdData) >= ntupSize)
+			{
+				*freeOffno = offno;
+				*freeNeighborOffno = neighborOffno;
+				return true;
+			}
+			else if (*nbuf != buf)
+				UnlockReleaseBuffer(*nbuf);
+		}
+	}
+
+	return false;
+}
+
+/*
+ * Add a new page
+ */
+static void
+HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page)
+{
+	/* Add a new page */
+	LockRelationForExtension(index, ExclusiveLock);
+	*nbuf = HnswNewBuffer(index, MAIN_FORKNUM);
+	UnlockRelationForExtension(index, ExclusiveLock);
+
+	/* Init new page */
+	*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
+	HnswInitPage(*nbuf, *npage);
+
+	/* Update previous buffer */
+	HnswPageGetOpaque(page)->nextblkno = BufferGetBlockNumber(*nbuf);
+}
+
+/*
+ * Add to element and neighbor pages
+ */
+static void
+WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage)
+{
+	Buffer		buf;
+	Page		page;
+	GenericXLogState *state;
+	Size		etupSize;
+	Size		ntupSize;
+	Size		combinedSize;
+	Size		maxSize;
+	Size		minCombinedSize;
+	HnswElementTuple etup;
+	BlockNumber currentPage = insertPage;
+	int			dimensions = e->vec->dim;
+	HnswNeighborTuple ntup;
+	Buffer		nbuf;
+	Page		npage;
+	OffsetNumber freeOffno = InvalidOffsetNumber;
+	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
+	BlockNumber newInsertPage = InvalidBlockNumber;
+
+	/* Calculate sizes */
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
+	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
+	maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
+	minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
+
+	/* Prepare element tuple */
+	etup = palloc0(etupSize);
+	HnswSetElementTuple(etup, e);
+
+	/* Prepare neighbor tuple */
+	ntup = palloc0(ntupSize);
+	HnswSetNeighborTuple(ntup, e, m);
+
+	/* Find a page (or two if needed) to insert the tuples */
+	for (;;)
+	{
+		buf = ReadBuffer(index, currentPage);
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+
+		/* Keep track of first page where element at level 0 can fit */
+		if (!BlockNumberIsValid(newInsertPage) && PageGetFreeSpace(page) >= minCombinedSize)
+			newInsertPage = currentPage;
+
+		/* First, try the fastest path */
+		/* Space for both tuples on the current page */
+		/* This can split existing tuples in rare cases */
+		if (PageGetFreeSpace(page) >= combinedSize)
+		{
+			nbuf = buf;
+			npage = page;
+			break;
+		}
+
+		/* Next, try space from a deleted element */
+		if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
+		{
+			if (nbuf != buf)
+				npage = GenericXLogRegisterBuffer(state, nbuf, 0);
+
+			break;
+		}
+
+		/* Finally, try space for element only if last page */
+		/* Skip if both tuples can fit on the same page */
+		if (combinedSize > maxSize && PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
+		{
+			HnswInsertAppendPage(index, &nbuf, &npage, state, page);
+			break;
+		}
+
+		currentPage = HnswPageGetOpaque(page)->nextblkno;
+
+		if (BlockNumberIsValid(currentPage))
+		{
+			/* Move to next page */
+			GenericXLogAbort(state);
+			UnlockReleaseBuffer(buf);
+		}
+		else
+		{
+			Buffer		newbuf;
+			Page		newpage;
+
+			HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
+
+			/* Commit */
+			MarkBufferDirty(newbuf);
+			MarkBufferDirty(buf);
+			GenericXLogFinish(state);
+
+			/* Unlock previous buffer */
+			UnlockReleaseBuffer(buf);
+
+			/* Prepare new buffer */
+			state = GenericXLogStart(index);
+			buf = newbuf;
+			page = GenericXLogRegisterBuffer(state, buf, 0);
+
+			/* Create new page for neighbors if needed */
+			if (PageGetFreeSpace(page) < combinedSize)
+				HnswInsertAppendPage(index, &nbuf, &npage, state, page);
+			else
+			{
+				nbuf = buf;
+				npage = page;
+			}
+
+			break;
+		}
+	}
+
+	e->blkno = BufferGetBlockNumber(buf);
+	e->neighborPage = BufferGetBlockNumber(nbuf);
+
+	/* Added tuple to new page if newInsertPage is not set */
+	/* So can set to neighbor page instead of element page */
+	if (!BlockNumberIsValid(newInsertPage))
+		newInsertPage = e->neighborPage;
+
+	if (OffsetNumberIsValid(freeOffno))
+	{
+		e->offno = freeOffno;
+		e->neighborOffno = freeNeighborOffno;
+	}
+	else
+	{
+		e->offno = OffsetNumberNext(PageGetMaxOffsetNumber(page));
+		if (nbuf == buf)
+			e->neighborOffno = OffsetNumberNext(e->offno);
+		else
+			e->neighborOffno = FirstOffsetNumber;
+	}
+
+	ItemPointerSet(&etup->neighbortid, e->neighborPage, e->neighborOffno);
+
+	/* Add element and neighbors */
+	if (OffsetNumberIsValid(freeOffno))
+	{
+		if (!PageIndexTupleOverwrite(page, e->offno, (Item) etup, etupSize))
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+		if (!PageIndexTupleOverwrite(npage, e->neighborOffno, (Item) ntup, ntupSize))
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+	}
+	else
+	{
+		if (PageAddItem(page, (Item) etup, etupSize, InvalidOffsetNumber, false, false) != e->offno)
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+		if (PageAddItem(npage, (Item) ntup, ntupSize, InvalidOffsetNumber, false, false) != e->neighborOffno)
+			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+	}
+
+	/* Commit */
+	MarkBufferDirty(buf);
+	if (nbuf != buf)
+		MarkBufferDirty(nbuf);
+	GenericXLogFinish(state);
+	UnlockReleaseBuffer(buf);
+	if (nbuf != buf)
+		UnlockReleaseBuffer(nbuf);
+
+	/* Update the insert page */
+	if (BlockNumberIsValid(newInsertPage) && newInsertPage != insertPage)
+		*updatedInsertPage = newInsertPage;
+}
+
+/*
+ * Check if connection already exists
+ */
+static bool
+ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
+{
+	for (int i = 0; i < lm; i++)
+	{
+		ItemPointer indextid = &ntup->indextids[startIdx + i];
+
+		if (!ItemPointerIsValid(indextid))
+			break;
+
+		if (ItemPointerGetBlockNumber(indextid) == e->blkno && ItemPointerGetOffsetNumber(indextid) == e->offno)
+			return true;
+	}
+
+	return false;
+}
+
+/*
+ * Update neighbors
+ */
+void
+HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting)
+{
+	for (int lc = e->level; lc >= 0; lc--)
+	{
+		int			lm = HnswGetLayerM(m, lc);
+		HnswNeighborArray *neighbors = &e->neighbors[lc];
+
+		for (int i = 0; i < neighbors->length; i++)
+		{
+			HnswCandidate *hc = &neighbors->items[i];
+			Buffer		buf;
+			Page		page;
+			GenericXLogState *state;
+			ItemId		itemid;
+			HnswNeighborTuple ntup;
+			Size		ntupSize;
+			int			idx = -1;
+			int			startIdx;
+			OffsetNumber offno = hc->element->neighborOffno;
+
+			/* Get latest neighbors since they may have changed */
+			/* Do not lock yet since selecting neighbors can take time */
+			HnswLoadNeighbors(hc->element, index, m);
+
+			/*
+			 * Could improve performance for vacuuming by checking neighbors
+			 * against list of elements being deleted to find index. It's
+			 * important to exclude already deleted elements for this since
+			 * they can be replaced at any time.
+			 */
+
+			/* Select neighbors */
+			HnswUpdateConnection(e, hc, lm, lc, &idx, index, procinfo, collation);
+
+			/* New element was not selected as a neighbor */
+			if (idx == -1)
+				continue;
+
+			/* Register page */
+			buf = ReadBuffer(index, hc->element->neighborPage);
+			LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+			state = GenericXLogStart(index);
+			page = GenericXLogRegisterBuffer(state, buf, 0);
+
+			/* Get tuple */
+			itemid = PageGetItemId(page, offno);
+			ntup = (HnswNeighborTuple) PageGetItem(page, itemid);
+			ntupSize = ItemIdGetLength(itemid);
+
+			/* Calculate index for update */
+			startIdx = (hc->element->level - lc) * m;
+
+			/* Check for existing connection */
+			if (checkExisting && ConnectionExists(e, ntup, startIdx, lm))
+				idx = -1;
+			else if (idx == -2)
+			{
+				/* Find free offset if still exists */
+				/* TODO Retry updating connections if not */
+				for (int j = 0; j < lm; j++)
+				{
+					if (!ItemPointerIsValid(&ntup->indextids[startIdx + j]))
+					{
+						idx = startIdx + j;
+						break;
+					}
+				}
+			}
+			else
+				idx += startIdx;
+
+			/* Make robust to issues */
+			if (idx >= 0 && idx < ntup->count)
+			{
+				ItemPointer indextid = &ntup->indextids[idx];
+
+				/* Update neighbor */
+				ItemPointerSet(indextid, e->blkno, e->offno);
+
+				/* Overwrite tuple */
+				if (!PageIndexTupleOverwrite(page, offno, (Item) ntup, ntupSize))
+					elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+				/* Commit */
+				MarkBufferDirty(buf);
+				GenericXLogFinish(state);
+			}
+			else
+				GenericXLogAbort(state);
+
+			UnlockReleaseBuffer(buf);
+		}
+	}
+}
+
+/*
+ * Add a heap TID to an existing element
+ */
+static bool
+HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
+{
+	Buffer		buf;
+	Page		page;
+	GenericXLogState *state;
+	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
+	HnswElementTuple etup;
+	int			i;
+
+	/* Read page */
+	buf = ReadBuffer(index, dup->blkno);
+	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+	state = GenericXLogStart(index);
+	page = GenericXLogRegisterBuffer(state, buf, 0);
+
+	/* Find space */
+	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
+	for (i = 0; i < HNSW_HEAPTIDS; i++)
+	{
+		if (!ItemPointerIsValid(&etup->heaptids[i]))
+			break;
+	}
+
+	/* Either being deleted or we lost our chance to another backend */
+	if (i == 0 || i == HNSW_HEAPTIDS)
+	{
+		GenericXLogAbort(state);
+		UnlockReleaseBuffer(buf);
+		return false;
+	}
+
+	/* Add heap TID */
+	etup->heaptids[i] = *((ItemPointer) linitial(element->heaptids));
+
+	/* Overwrite tuple */
+	if (!PageIndexTupleOverwrite(page, dup->offno, (Item) etup, etupSize))
+		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+	/* Commit */
+	MarkBufferDirty(buf);
+	GenericXLogFinish(state);
+	UnlockReleaseBuffer(buf);
+
+	return true;
+}
+
+/*
+ * Write changes to disk
+ */
+static void
+WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement dup, HnswElement entryPoint)
+{
+	BlockNumber newInsertPage = InvalidBlockNumber;
+
+	/* Try to add to existing page */
+	if (dup != NULL)
+	{
+		if (HnswAddDuplicate(index, element, dup))
+			return;
+	}
+
+	/* Write element and neighbor tuples */
+	WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage);
+
+	/* Update insert page if needed */
+	if (BlockNumberIsValid(newInsertPage))
+		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM);
+
+	/* Update neighbors */
+	HnswUpdateNeighborPages(index, procinfo, collation, element, m, false);
+
+	/* Update metapage if needed */
+	if (entryPoint == NULL || element->level > entryPoint->level)
+		HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
+}
+
+/*
+ * Insert a tuple into the index
+ */
+bool
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
+{
+	Datum		value;
+	FmgrInfo   *normprocinfo;
+	HnswElement entryPoint;
+	HnswElement element;
+	int			m;
+	int			efConstruction = HnswGetEfConstruction(index);
+	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	Oid			collation = index->rd_indcollation[0];
+	HnswElement dup;
+	LOCKMODE	lockmode = ShareLock;
+
+	/* Detoast once for all calls */
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Normalize if needed */
+	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	if (normprocinfo != NULL)
+	{
+		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
+			return false;
+	}
+
+	/*
+	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
+	 * before repairing graph. Use a page lock so it does not interfere with
+	 * buffer lock (or reads when vacuuming).
+	 */
+	LockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+	/* Get m and entry point */
+	HnswGetMetaPageInfo(index, &m, &entryPoint);
+
+	/* Create an element */
+	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
+	element->vec = DatumGetVector(value);
+
+	/* Prevent concurrent inserts when likely updating entry point */
+	if (entryPoint == NULL || element->level > entryPoint->level)
+	{
+		/* Release shared lock */
+		UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+		/* Get exclusive lock */
+		lockmode = ExclusiveLock;
+		LockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+		/* Get latest entry point after lock is acquired */
+		entryPoint = HnswGetEntryPoint(index);
+	}
+
+	/* Insert element in graph */
+	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, false);
+
+	/* Look for duplicate */
+	dup = HnswFindDuplicate(element);
+
+	/* Write to disk */
+	WriteElement(index, procinfo, collation, element, m, efConstruction, dup, entryPoint);
+
+	/* Release lock */
+	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+	return true;
+}
+
+/*
+ * Insert a tuple into the index
+ */
+bool
+hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
+		   Relation heap, IndexUniqueCheck checkUnique
+#if PG_VERSION_NUM >= 140000
+		   ,bool indexUnchanged
+#endif
+		   ,IndexInfo *indexInfo
+)
+{
+	MemoryContext oldCtx;
+	MemoryContext insertCtx;
+
+	/* Skip nulls */
+	if (isnull[0])
+		return false;
+
+	/* Create memory context */
+	insertCtx = AllocSetContextCreate(CurrentMemoryContext,
+									  "Hnsw insert temporary context",
+									  ALLOCSET_DEFAULT_SIZES);
+	oldCtx = MemoryContextSwitchTo(insertCtx);
+
+	/* Insert tuple */
+	HnswInsertTuple(index, values, isnull, heap_tid, heap);
+
+	/* Delete memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextDelete(insertCtx);
+
+	return false;
+}

src/hnswscan.c

@@ -0,0 +1,233 @@
+#include "postgres.h"
+
+#include "access/relscan.h"
+#include "hnsw.h"
+#include "pgstat.h"
+#include "storage/bufmgr.h"
+#include "storage/lmgr.h"
+#include "utils/memutils.h"
+
+/*
+ * Algorithm 5 from paper
+ */
+static List *
+GetScanItems(IndexScanDesc scan, Datum q)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	Relation	index = scan->indexRelation;
+	FmgrInfo   *procinfo = so->procinfo;
+	Oid			collation = so->collation;
+	List	   *ep;
+	List	   *w;
+	int			m;
+	HnswElement entryPoint;
+
+	/* Get m and entry point */
+	HnswGetMetaPageInfo(index, &m, &entryPoint);
+
+	if (entryPoint == NULL)
+		return NIL;
+
+	ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
+
+	for (int lc = entryPoint->level; lc >= 1; lc--)
+	{
+		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
+		ep = w;
+	}
+
+	return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
+}
+
+/*
+ * Get dimensions from metapage
+ */
+static int
+GetDimensions(Relation index)
+{
+	Buffer		buf;
+	Page		page;
+	HnswMetaPage metap;
+	int			dimensions;
+
+	buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	metap = HnswPageGetMeta(page);
+
+	dimensions = metap->dimensions;
+
+	UnlockReleaseBuffer(buf);
+
+	return dimensions;
+}
+
+/*
+ * Get scan value
+ */
+static Datum
+GetScanValue(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	Datum		value;
+
+	if (scan->orderByData->sk_flags & SK_ISNULL)
+		value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
+	else
+	{
+		value = scan->orderByData->sk_argument;
+
+		/* Value should not be compressed or toasted */
+		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
+
+		/* Fine if normalization fails */
+		if (so->normprocinfo != NULL)
+			HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
+	}
+
+	return value;
+}
+
+/*
+ * Prepare for an index scan
+ */
+IndexScanDesc
+hnswbeginscan(Relation index, int nkeys, int norderbys)
+{
+	IndexScanDesc scan;
+	HnswScanOpaque so;
+
+	scan = RelationGetIndexScan(index, nkeys, norderbys);
+
+	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
+	so->first = true;
+	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+									   "Hnsw scan temporary context",
+									   ALLOCSET_DEFAULT_SIZES);
+
+	/* Set support functions */
+	so->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	so->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	so->collation = index->rd_indcollation[0];
+
+	scan->opaque = so;
+
+	/*
+	 * Get a shared lock. This allows vacuum to ensure no in-flight scans
+	 * before marking tuples as deleted.
+	 */
+	LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+	return scan;
+}
+
+/*
+ * Start or restart an index scan
+ */
+void
+hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int norderbys)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+
+	so->first = true;
+	MemoryContextReset(so->tmpCtx);
+
+	if (keys && scan->numberOfKeys > 0)
+		memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));
+
+	if (orderbys && scan->numberOfOrderBys > 0)
+		memmove(scan->orderByData, orderbys, scan->numberOfOrderBys * sizeof(ScanKeyData));
+}
+
+/*
+ * Fetch the next tuple in the given scan
+ */
+bool
+hnswgettuple(IndexScanDesc scan, ScanDirection dir)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	MemoryContext oldCtx = MemoryContextSwitchTo(so->tmpCtx);
+
+	/*
+	 * Index can be used to scan backward, but Postgres doesn't support
+	 * backward scan on operators
+	 */
+	Assert(ScanDirectionIsForward(dir));
+
+	if (so->first)
+	{
+		Datum		value;
+
+		/* Count index scan for stats */
+		pgstat_count_index_scan(scan->indexRelation);
+
+		/* Safety check */
+		if (scan->orderByData == NULL)
+			elog(ERROR, "cannot scan hnsw index without order");
+
+		/* Get scan value */
+		value = GetScanValue(scan);
+
+		so->w = GetScanItems(scan, value);
+
+		so->first = false;
+	}
+
+	while (list_length(so->w) > 0)
+	{
+		HnswCandidate *hc = llast(so->w);
+		ItemPointer heaptid;
+
+		/* Move to next element if no valid heap TIDs */
+		if (list_length(hc->element->heaptids) == 0)
+		{
+			so->w = list_delete_last(so->w);
+			continue;
+		}
+
+		heaptid = llast(hc->element->heaptids);
+
+		hc->element->heaptids = list_delete_last(hc->element->heaptids);
+
+		MemoryContextSwitchTo(oldCtx);
+
+#if PG_VERSION_NUM >= 120000
+		scan->xs_heaptid = *heaptid;
+#else
+		scan->xs_ctup.t_self = *heaptid;
+#endif
+
+		/*
+		 * Typically, an index scan must maintain a pin on the index page
+		 * holding the item last returned by amgettuple. However, this is not
+		 * needed with the current vacuum strategy, which ensures scans do not
+		 * visit tuples in danger of being marked as deleted.
+		 *
+		 * https://www.postgresql.org/docs/current/index-locking.html
+		 */
+
+		scan->xs_recheckorderby = false;
+		return true;
+	}
+
+	MemoryContextSwitchTo(oldCtx);
+	return false;
+}
+
+/*
+ * End a scan and release resources
+ */
+void
+hnswendscan(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+
+	/* Release shared lock */
+	UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
+
+	MemoryContextDelete(so->tmpCtx);
+
+	pfree(so);
+	scan->opaque = NULL;
+}

src/hnswvacuum.c

@@ -0,0 +1,667 @@
+#include "postgres.h"
+
+#include <math.h>
+
+#include "commands/vacuum.h"
+#include "hnsw.h"
+#include "storage/bufmgr.h"
+#include "storage/lmgr.h"
+#include "utils/memutils.h"
+
+/*
+ * Check if deleted list contains an index TID
+ */
+static bool
+DeletedContains(HTAB *deleted, ItemPointer indextid)
+{
+	bool		found;
+
+	hash_search(deleted, indextid, HASH_FIND, &found);
+	return found;
+}
+
+/*
+ * Remove deleted heap TIDs
+ *
+ * OK to remove for entry point, since always considered for searches and inserts
+ */
+static void
+RemoveHeapTids(HnswVacuumState * vacuumstate)
+{
+	BlockNumber blkno = HNSW_HEAD_BLKNO;
+	HnswElement highestPoint = &vacuumstate->highestPoint;
+	Relation	index = vacuumstate->index;
+	BufferAccessStrategy bas = vacuumstate->bas;
+	HnswElement entryPoint = HnswGetEntryPoint(vacuumstate->index);
+	IndexBulkDeleteResult *stats = vacuumstate->stats;
+
+	/* Store separately since highestPoint.level is uint8 */
+	int			highestLevel = -1;
+
+	/* Initialize highest point */
+	highestPoint->blkno = InvalidBlockNumber;
+	highestPoint->offno = InvalidOffsetNumber;
+
+	while (BlockNumberIsValid(blkno))
+	{
+		Buffer		buf;
+		Page		page;
+		GenericXLogState *state;
+		OffsetNumber offno;
+		OffsetNumber maxoffno;
+		bool		updated = false;
+
+		vacuum_delay_point();
+
+		buf = ReadBufferExtended(index, MAIN_FORKNUM, blkno, RBM_NORMAL, bas);
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+		maxoffno = PageGetMaxOffsetNumber(page);
+
+		/* Iterate over nodes */
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		{
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			int			idx = 0;
+			bool		itemUpdated = false;
+
+			/* Skip neighbor tuples */
+			if (!HnswIsElementTuple(etup))
+				continue;
+
+			if (ItemPointerIsValid(&etup->heaptids[0]))
+			{
+				for (int i = 0; i < HNSW_HEAPTIDS; i++)
+				{
+					/* Stop at first unused */
+					if (!ItemPointerIsValid(&etup->heaptids[i]))
+						break;
+
+					if (vacuumstate->callback(&etup->heaptids[i], vacuumstate->callback_state))
+					{
+						itemUpdated = true;
+						stats->tuples_removed++;
+					}
+					else
+					{
+						/* Move to front of list */
+						etup->heaptids[idx++] = etup->heaptids[i];
+						stats->num_index_tuples++;
+					}
+				}
+
+				if (itemUpdated)
+				{
+					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+
+					/* Mark rest as invalid */
+					for (int i = idx; i < HNSW_HEAPTIDS; i++)
+						ItemPointerSetInvalid(&etup->heaptids[i]);
+
+					if (!PageIndexTupleOverwrite(page, offno, (Item) etup, etupSize))
+						elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+					updated = true;
+				}
+			}
+
+			if (!ItemPointerIsValid(&etup->heaptids[0]))
+			{
+				ItemPointerData ip;
+
+				/* Add to deleted list */
+				ItemPointerSet(&ip, blkno, offno);
+
+				(void) hash_search(vacuumstate->deleted, &ip, HASH_ENTER, NULL);
+			}
+			else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
+			{
+				/* Keep track of highest non-entry point */
+				highestPoint->blkno = blkno;
+				highestPoint->offno = offno;
+				highestPoint->level = etup->level;
+				highestLevel = etup->level;
+			}
+		}
+
+		blkno = HnswPageGetOpaque(page)->nextblkno;
+
+		if (updated)
+		{
+			MarkBufferDirty(buf);
+			GenericXLogFinish(state);
+		}
+		else
+			GenericXLogAbort(state);
+
+		UnlockReleaseBuffer(buf);
+	}
+}
+
+/*
+ * Check for deleted neighbors
+ */
+static bool
+NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
+{
+	Relation	index = vacuumstate->index;
+	BufferAccessStrategy bas = vacuumstate->bas;
+	Buffer		buf;
+	Page		page;
+	HnswNeighborTuple ntup;
+	bool		needsUpdated = false;
+
+	buf = ReadBufferExtended(index, MAIN_FORKNUM, element->neighborPage, RBM_NORMAL, bas);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
+
+	Assert(HnswIsNeighborTuple(ntup));
+
+	/* Check neighbors */
+	for (int i = 0; i < ntup->count; i++)
+	{
+		ItemPointer indextid = &ntup->indextids[i];
+
+		if (!ItemPointerIsValid(indextid))
+			continue;
+
+		/* Check if in deleted list */
+		if (DeletedContains(vacuumstate->deleted, indextid))
+		{
+			needsUpdated = true;
+			break;
+		}
+	}
+
+	/* Also update if layer 0 is not full */
+	/* This could indicate too many candidates being deleted during insert */
+	if (!needsUpdated)
+		needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
+
+	UnlockReleaseBuffer(buf);
+
+	return needsUpdated;
+}
+
+/*
+ * Repair graph for a single element
+ */
+static void
+RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
+{
+	Relation	index = vacuumstate->index;
+	Buffer		buf;
+	Page		page;
+	GenericXLogState *state;
+	int			m = vacuumstate->m;
+	int			efConstruction = vacuumstate->efConstruction;
+	FmgrInfo   *procinfo = vacuumstate->procinfo;
+	Oid			collation = vacuumstate->collation;
+	BufferAccessStrategy bas = vacuumstate->bas;
+	HnswNeighborTuple ntup = vacuumstate->ntup;
+	Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
+
+	/* Skip if element is entry point */
+	if (entryPoint != NULL && element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
+		return;
+
+	/* Init fields */
+	HnswInitNeighbors(element, m);
+	element->heaptids = NIL;
+
+	/* Add element to graph, skipping itself */
+	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, true);
+
+	/* Update neighbor tuple */
+	/* Do this before getting page to minimize locking */
+	HnswSetNeighborTuple(ntup, element, m);
+
+	/* Get neighbor page */
+	buf = ReadBufferExtended(index, MAIN_FORKNUM, element->neighborPage, RBM_NORMAL, bas);
+	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+	state = GenericXLogStart(index);
+	page = GenericXLogRegisterBuffer(state, buf, 0);
+
+	/* Overwrite tuple */
+	if (!PageIndexTupleOverwrite(page, element->neighborOffno, (Item) ntup, ntupSize))
+		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+	/* Commit */
+	MarkBufferDirty(buf);
+	GenericXLogFinish(state);
+	UnlockReleaseBuffer(buf);
+
+	/* Update neighbors */
+	HnswUpdateNeighborPages(index, procinfo, collation, element, m, true);
+}
+
+/*
+ * Repair graph entry point
+ */
+static void
+RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
+{
+	Relation	index = vacuumstate->index;
+	HnswElement highestPoint = &vacuumstate->highestPoint;
+	HnswElement entryPoint;
+	MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
+
+	if (!BlockNumberIsValid(highestPoint->blkno))
+		highestPoint = NULL;
+
+	/*
+	 * Repair graph for highest non-entry point. Highest point may be outdated
+	 * due to inserts that happen during and after RemoveHeapTids.
+	 */
+	if (highestPoint != NULL)
+	{
+		/* Get a shared lock */
+		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
+
+		/* Load element */
+		HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
+
+		/* Repair if needed */
+		if (NeedsUpdated(vacuumstate, highestPoint))
+			RepairGraphElement(vacuumstate, highestPoint, HnswGetEntryPoint(index));
+
+		/* Release lock */
+		UnlockPage(index, HNSW_UPDATE_LOCK, ShareLock);
+	}
+
+	/* Prevent concurrent inserts when possibly updating entry point */
+	LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+
+	/* Get latest entry point */
+	entryPoint = HnswGetEntryPoint(index);
+
+	if (entryPoint != NULL)
+	{
+		ItemPointerData epData;
+
+		ItemPointerSet(&epData, entryPoint->blkno, entryPoint->offno);
+
+		if (DeletedContains(vacuumstate->deleted, &epData))
+		{
+			/*
+			 * Replace the entry point with the highest point. If highest
+			 * point is outdated and empty, the entry point will be empty
+			 * until an element is repaired.
+			 */
+			HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, highestPoint, InvalidBlockNumber, MAIN_FORKNUM);
+		}
+		else
+		{
+			/*
+			 * Repair the entry point with the highest point. If highest point
+			 * is outdated, this can remove connections at higher levels in
+			 * the graph until they are repaired, but this should be fine.
+			 */
+			HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
+
+			if (NeedsUpdated(vacuumstate, entryPoint))
+			{
+				/* Reset neighbors from previous update */
+				if (highestPoint != NULL)
+					highestPoint->neighbors = NULL;
+
+				RepairGraphElement(vacuumstate, entryPoint, highestPoint);
+			}
+		}
+	}
+
+	/* Release lock */
+	UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+
+	/* Reset memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextReset(vacuumstate->tmpCtx);
+}
+
+/*
+ * Repair graph for all elements
+ */
+static void
+RepairGraph(HnswVacuumState * vacuumstate)
+{
+	Relation	index = vacuumstate->index;
+	BufferAccessStrategy bas = vacuumstate->bas;
+	BlockNumber blkno = HNSW_HEAD_BLKNO;
+
+	/*
+	 * Wait for inserts to complete. Inserts before this point may have
+	 * neighbors about to be deleted. Inserts after this point will not.
+	 */
+	LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+	UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+
+	/* Repair entry point first */
+	RepairGraphEntryPoint(vacuumstate);
+
+	while (BlockNumberIsValid(blkno))
+	{
+		Buffer		buf;
+		Page		page;
+		OffsetNumber offno;
+		OffsetNumber maxoffno;
+		List	   *elements = NIL;
+		ListCell   *lc2;
+		MemoryContext oldCtx;
+
+		vacuum_delay_point();
+
+		oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
+
+		buf = ReadBufferExtended(index, MAIN_FORKNUM, blkno, RBM_NORMAL, bas);
+		LockBuffer(buf, BUFFER_LOCK_SHARE);
+		page = BufferGetPage(buf);
+		maxoffno = PageGetMaxOffsetNumber(page);
+
+		/* Load items into memory to minimize locking */
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		{
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			HnswElement element;
+
+			/* Skip neighbor tuples */
+			if (!HnswIsElementTuple(etup))
+				continue;
+
+			/* Skip updating neighbors if being deleted */
+			if (!ItemPointerIsValid(&etup->heaptids[0]))
+				continue;
+
+			/* Create an element */
+			element = HnswInitElementFromBlock(blkno, offno);
+			HnswLoadElementFromTuple(element, etup, false, true);
+
+			elements = lappend(elements, element);
+		}
+
+		blkno = HnswPageGetOpaque(page)->nextblkno;
+
+		UnlockReleaseBuffer(buf);
+
+		/* Update neighbor pages */
+		foreach(lc2, elements)
+		{
+			HnswElement element = (HnswElement) lfirst(lc2);
+			HnswElement entryPoint;
+			LOCKMODE	lockmode = ShareLock;
+
+			/* Check if any neighbors point to deleted values */
+			if (!NeedsUpdated(vacuumstate, element))
+				continue;
+
+			/* Get a shared lock */
+			LockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+			/* Refresh entry point for each element */
+			entryPoint = HnswGetEntryPoint(index);
+
+			/* Prevent concurrent inserts when likely updating entry point */
+			if (entryPoint == NULL || element->level > entryPoint->level)
+			{
+				/* Release shared lock */
+				UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+				/* Get exclusive lock */
+				lockmode = ExclusiveLock;
+				LockPage(index, HNSW_UPDATE_LOCK, lockmode);
+
+				/* Get latest entry point after lock is acquired */
+				entryPoint = HnswGetEntryPoint(index);
+			}
+
+			/* Repair connections */
+			RepairGraphElement(vacuumstate, element, entryPoint);
+
+			/*
+			 * Update metapage if needed. Should only happen if entry point
+			 * was replaced and highest point was outdated.
+			 */
+			if (entryPoint == NULL || element->level > entryPoint->level)
+				HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
+
+			/* Release lock */
+			UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
+		}
+
+		/* Reset memory context */
+		MemoryContextSwitchTo(oldCtx);
+		MemoryContextReset(vacuumstate->tmpCtx);
+	}
+}
+
+/*
+ * Mark items as deleted
+ */
+static void
+MarkDeleted(HnswVacuumState * vacuumstate)
+{
+	BlockNumber blkno = HNSW_HEAD_BLKNO;
+	BlockNumber insertPage = InvalidBlockNumber;
+	Relation	index = vacuumstate->index;
+	BufferAccessStrategy bas = vacuumstate->bas;
+
+	/*
+	 * Wait for index scans to complete. Scans before this point may contain
+	 * tuples about to be deleted. Scans after this point will not, since the
+	 * graph has been repaired.
+	 */
+	LockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
+	UnlockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
+
+	while (BlockNumberIsValid(blkno))
+	{
+		Buffer		buf;
+		Page		page;
+		GenericXLogState *state;
+		OffsetNumber offno;
+		OffsetNumber maxoffno;
+
+		vacuum_delay_point();
+
+		buf = ReadBufferExtended(index, MAIN_FORKNUM, blkno, RBM_NORMAL, bas);
+
+		/*
+		 * ambulkdelete cannot delete entries from pages that are pinned by
+		 * other backends
+		 *
+		 * https://www.postgresql.org/docs/current/index-locking.html
+		 */
+		LockBufferForCleanup(buf);
+
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+		maxoffno = PageGetMaxOffsetNumber(page);
+
+		/* Update element and neighbors together */
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		{
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			HnswNeighborTuple ntup;
+			Size		etupSize;
+			Size		ntupSize;
+			Buffer		nbuf;
+			Page		npage;
+			BlockNumber neighborPage;
+			OffsetNumber neighborOffno;
+
+			/* Skip neighbor tuples */
+			if (!HnswIsElementTuple(etup))
+				continue;
+
+			/* Skip deleted tuples */
+			if (etup->deleted)
+			{
+				/* Set to first free page */
+				if (!BlockNumberIsValid(insertPage))
+					insertPage = blkno;
+
+				continue;
+			}
+
+			/* Skip live tuples */
+			if (ItemPointerIsValid(&etup->heaptids[0]))
+				continue;
+
+			/* Calculate sizes */
+			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
+
+			/* Get neighbor page */
+			neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
+			neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
+
+			if (neighborPage == blkno)
+			{
+				nbuf = buf;
+				npage = page;
+			}
+			else
+			{
+				nbuf = ReadBufferExtended(index, MAIN_FORKNUM, neighborPage, RBM_NORMAL, bas);
+				LockBuffer(nbuf, BUFFER_LOCK_EXCLUSIVE);
+				npage = GenericXLogRegisterBuffer(state, nbuf, 0);
+			}
+
+			ntup = (HnswNeighborTuple) PageGetItem(npage, PageGetItemId(npage, neighborOffno));
+
+			/* Overwrite element */
+			etup->deleted = 1;
+			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
+
+			/* Overwrite neighbors */
+			for (int i = 0; i < ntup->count; i++)
+				ItemPointerSetInvalid(&ntup->indextids[i]);
+
+			/* Overwrite element tuple */
+			if (!PageIndexTupleOverwrite(page, offno, (Item) etup, etupSize))
+				elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+			/* Overwrite neighbor tuple */
+			if (!PageIndexTupleOverwrite(npage, neighborOffno, (Item) ntup, ntupSize))
+				elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+
+			/* Commit */
+			MarkBufferDirty(buf);
+			if (nbuf != buf)
+				MarkBufferDirty(nbuf);
+			GenericXLogFinish(state);
+			if (nbuf != buf)
+				UnlockReleaseBuffer(nbuf);
+
+			/* Set to first free page */
+			if (!BlockNumberIsValid(insertPage))
+				insertPage = blkno;
+
+			/* Prepare new xlog */
+			state = GenericXLogStart(index);
+			page = GenericXLogRegisterBuffer(state, buf, 0);
+		}
+
+		blkno = HnswPageGetOpaque(page)->nextblkno;
+
+		GenericXLogAbort(state);
+		UnlockReleaseBuffer(buf);
+	}
+
+	/* Update insert page last, after everything has been marked as deleted */
+	HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM);
+}
+
+/*
+ * Initialize the vacuum state
+ */
+static void
+InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkDeleteResult *stats, IndexBulkDeleteCallback callback, void *callback_state)
+{
+	Relation	index = info->index;
+	HASHCTL		hash_ctl;
+
+	if (stats == NULL)
+		stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
+
+	vacuumstate->index = index;
+	vacuumstate->stats = stats;
+	vacuumstate->callback = callback;
+	vacuumstate->callback_state = callback_state;
+	vacuumstate->efConstruction = HnswGetEfConstruction(index);
+	vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
+	vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	vacuumstate->collation = index->rd_indcollation[0];
+	vacuumstate->ntup = palloc0(BLCKSZ);
+	vacuumstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+												"Hnsw vacuum temporary context",
+												ALLOCSET_DEFAULT_SIZES);
+
+	/* Get m from metapage */
+	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
+
+	/* Create hash table */
+	hash_ctl.keysize = sizeof(ItemPointerData);
+	hash_ctl.entrysize = sizeof(ItemPointerData);
+	hash_ctl.hcxt = CurrentMemoryContext;
+	vacuumstate->deleted = hash_create("hnswbulkdelete indextids", 256, &hash_ctl, HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
+}
+
+/*
+ * Free resources
+ */
+static void
+FreeVacuumState(HnswVacuumState * vacuumstate)
+{
+	hash_destroy(vacuumstate->deleted);
+	FreeAccessStrategy(vacuumstate->bas);
+	pfree(vacuumstate->ntup);
+	MemoryContextDelete(vacuumstate->tmpCtx);
+}
+
+/*
+ * Bulk delete tuples from the index
+ */
+IndexBulkDeleteResult *
+hnswbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
+			   IndexBulkDeleteCallback callback, void *callback_state)
+{
+	HnswVacuumState vacuumstate;
+
+	InitVacuumState(&vacuumstate, info, stats, callback, callback_state);
+
+	/* Pass 1: Remove heap TIDs */
+	RemoveHeapTids(&vacuumstate);
+
+	/* Pass 2: Repair graph */
+	RepairGraph(&vacuumstate);
+
+	/* Pass 3: Mark as deleted */
+	MarkDeleted(&vacuumstate);
+
+	FreeVacuumState(&vacuumstate);
+
+	return vacuumstate.stats;
+}
+
+/*
+ * Clean up after a VACUUM operation
+ */
+IndexBulkDeleteResult *
+hnswvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
+{
+	Relation	rel = info->index;
+
+	if (info->analyze_only)
+		return stats;
+
+	/* stats is NULL if ambulkdelete not called */
+	/* OK to return NULL if index not changed */
+	if (stats == NULL)
+		return NULL;
+
+	stats->num_pages = RelationGetNumberOfBlocks(rel);
+
+	return stats;
+}

src/ivfflat.c

@@ -1,22 +1,30 @@
 #include "postgres.h"
 
+#include <float.h>
+
 #include "access/amapi.h"
 #include "commands/vacuum.h"
 #include "ivfflat.h"
 #include "utils/guc.h"
 #include "utils/selfuncs.h"
+#include "utils/spccache.h"
 
+#if PG_VERSION_NUM >= 120000
+#include "commands/progress.h"
+#endif
+
+int			ivfflat_probes;
 static relopt_kind ivfflat_relopt_kind;
 
 /*
  * Initialize index options and variables
  */
 void
-_PG_init(void)
+IvfflatInit(void)
 {
 	ivfflat_relopt_kind = add_reloption_kind();
 	add_int_reloption(ivfflat_relopt_kind, "lists", "Number of inverted lists",
-					  IVFFLAT_DEFAULT_LISTS, 1, IVFFLAT_MAX_LISTS
+					  IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS
 #if PG_VERSION_NUM >= 130000
 					  ,AccessExclusiveLock
 #endif
@@ -24,58 +32,116 @@ _PG_init(void)
 
 	DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
 							"Valid range is 1..lists.", &ivfflat_probes,
-							1, 1, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
+							IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
+/*
+ * Get the name of index build phase
+ */
+#if PG_VERSION_NUM >= 120000
+static char *
+ivfflatbuildphasename(int64 phasenum)
+{
+	switch (phasenum)
+	{
+		case PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE:
+			return "initializing";
+		case PROGRESS_IVFFLAT_PHASE_KMEANS:
+			return "performing k-means";
+		case PROGRESS_IVFFLAT_PHASE_ASSIGN:
+			return "assigning tuples";
+		case PROGRESS_IVFFLAT_PHASE_LOAD:
+			return "loading tuples";
+		default:
+			return NULL;
+	}
+}
+#endif
+
 /*
  * Estimate the cost of an index scan
  */
 static void
 ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 					Cost *indexStartupCost, Cost *indexTotalCost,
-					Selectivity *indexSelectivity, double *indexCorrelation
-#if PG_VERSION_NUM >= 100000
-					,double *indexPages
-#endif
-)
+					Selectivity *indexSelectivity, double *indexCorrelation,
+					double *indexPages)
 {
 	GenericCosts costs;
 	int			lists;
 	double		ratio;
-	Relation	indexRel;
-
+	double		spc_seq_page_cost;
+	Relation	index;
 #if PG_VERSION_NUM < 120000
 	List	   *qinfos;
-
-	qinfos = deconstruct_indexquals(path);
 #endif
 
+	/* Never use index without order */
+	if (path->indexorderbys == NULL)
+	{
+		*indexStartupCost = DBL_MAX;
+		*indexTotalCost = DBL_MAX;
+		*indexSelectivity = 0;
+		*indexCorrelation = 0;
+		*indexPages = 0;
+		return;
+	}
+
 	MemSet(&costs, 0, sizeof(costs));
 
+	index = index_open(path->indexinfo->indexoid, NoLock);
+	IvfflatGetMetaPageInfo(index, &lists, NULL);
+	index_close(index, NoLock);
+
+	/* Get the ratio of lists that we need to visit */
+	ratio = ((double) ivfflat_probes) / lists;
+	if (ratio > 1.0)
+		ratio = 1.0;
+
+	/*
+	 * This gives us the subset of tuples to visit. This value is passed into
+	 * the generic cost estimator to determine the number of pages to visit
+	 * during the index scan.
+	 */
+	costs.numIndexTuples = path->indexinfo->tuples * ratio;
+
 #if PG_VERSION_NUM >= 120000
 	genericcostestimate(root, path, loop_count, &costs);
 #else
+	qinfos = deconstruct_indexquals(path);
 	genericcostestimate(root, path, loop_count, qinfos, &costs);
 #endif
 
-	indexRel = index_open(path->indexinfo->indexoid, NoLock);
-	lists = IvfflatGetLists(indexRel);
-	index_close(indexRel, NoLock);
+	get_tablespace_page_costs(path->indexinfo->reltablespace, NULL, &spc_seq_page_cost);
 
-	ratio = ((double) ivfflat_probes) / lists;
-	if (ratio > 1)
-		ratio = 1;
+	/* Adjust cost if needed since TOAST not included in seq scan cost */
+	if (costs.numIndexPages > path->indexinfo->rel->pages && ratio < 0.5)
+	{
+		/* Change all page cost from random to sequential */
+		costs.indexTotalCost -= costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);
 
-	costs.indexTotalCost *= ratio;
+		/* Remove cost of extra pages */
+		costs.indexTotalCost -= (costs.numIndexPages - path->indexinfo->rel->pages) * spc_seq_page_cost;
+	}
+	else
+	{
+		/* Change some page cost from random to sequential */
+		costs.indexTotalCost -= 0.5 * costs.numIndexPages * (costs.spc_random_page_cost - spc_seq_page_cost);
+	}
 
-	/* Startup cost and total cost are same */
+	/*
+	 * If the list selectivity is lower than what is returned from the generic
+	 * cost estimator, use that.
+	 */
+	if (ratio < costs.indexSelectivity)
+		costs.indexSelectivity = ratio;
+
+	/* Use total cost since most work happens before first tuple is returned */
 	*indexStartupCost = costs.indexTotalCost;
 	*indexTotalCost = costs.indexTotalCost;
 	*indexSelectivity = costs.indexSelectivity;
 	*indexCorrelation = costs.indexCorrelation;
-#if PG_VERSION_NUM >= 100000
 	*indexPages = costs.numIndexPages;
-#endif
 }
 
 /*
@@ -116,7 +182,12 @@ ivfflatvalidate(Oid opclassoid)
 	return true;
 }
 
-PG_FUNCTION_INFO_V1(ivfflathandler);
+/*
+ * Define index handler
+ *
+ * See https://www.postgresql.org/docs/current/index-api.html
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(ivfflathandler);
 Datum
 ivfflathandler(PG_FUNCTION_ARGS)
 {
@@ -138,31 +209,31 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	amroutine->amstorage = false;
 	amroutine->amclusterable = false;
 	amroutine->ampredlocks = false;
-#if PG_VERSION_NUM >= 100000
 	amroutine->amcanparallel = false;
-#endif
-#if PG_VERSION_NUM >= 110000
 	amroutine->amcaninclude = false;
-#endif
 #if PG_VERSION_NUM >= 130000
 	amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
-	amroutine->amparallelvacuumoptions = VACUUM_OPTION_NO_PARALLEL; /* TODO support parallel */
+	amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
 #endif
 	amroutine->amkeytype = InvalidOid;
 
+	/* Interface functions */
 	amroutine->ambuild = ivfflatbuild;
 	amroutine->ambuildempty = ivfflatbuildempty;
 	amroutine->aminsert = ivfflatinsert;
 	amroutine->ambulkdelete = ivfflatbulkdelete;
 	amroutine->amvacuumcleanup = ivfflatvacuumcleanup;
-	amroutine->amcanreturn = NULL;
+	amroutine->amcanreturn = NULL;	/* tuple not included in heapsort */
 	amroutine->amcostestimate = ivfflatcostestimate;
 	amroutine->amoptions = ivfflatoptions;
 	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */
 #if PG_VERSION_NUM >= 120000
-	amroutine->ambuildphasename = NULL;
+	amroutine->ambuildphasename = ivfflatbuildphasename;
 #endif
 	amroutine->amvalidate = ivfflatvalidate;
+#if PG_VERSION_NUM >= 140000
+	amroutine->amadjustmembers = NULL;
+#endif
 	amroutine->ambeginscan = ivfflatbeginscan;
 	amroutine->amrescan = ivfflatrescan;
 	amroutine->amgettuple = ivfflatgettuple;
@@ -170,11 +241,11 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	amroutine->amendscan = ivfflatendscan;
 	amroutine->ammarkpos = NULL;
 	amroutine->amrestrpos = NULL;
-#if PG_VERSION_NUM >= 100000
+
+	/* Interface functions to support parallel index scans */
 	amroutine->amestimateparallelscan = NULL;
 	amroutine->aminitparallelscan = NULL;
 	amroutine->amparallelrescan = NULL;
-#endif
 
 	PG_RETURN_POINTER(amroutine);
 }

src/ivfflat.h

@@ -4,12 +4,28 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
+#include "access/parallel.h"
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
+#include "port.h"				/* for random() */
 #include "utils/sampling.h"
 #include "utils/tuplesort.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 150000
+#include "common/pg_prng.h"
+#endif
+
+#if PG_VERSION_NUM < 120000
+#include "access/relscan.h"
+#endif
+
+#ifdef IVFFLAT_BENCH
+#include "portability/instr_time.h"
+#endif
+
+#define IVFFLAT_MAX_DIM 2000
+
 /* Support functions */
 #define IVFFLAT_DISTANCE_PROC 1
 #define IVFFLAT_NORM_PROC 2
@@ -24,28 +40,58 @@
 #define IVFFLAT_METAPAGE_BLKNO	0
 #define IVFFLAT_HEAD_BLKNO		1	/* first list page */
 
+/* Must correspond to page numbers since page lock is used */
+#define IVFFLAT_SCAN_LOCK		0
+
+/* IVFFlat parameters */
 #define IVFFLAT_DEFAULT_LISTS	100
+#define IVFFLAT_MIN_LISTS		1
 #define IVFFLAT_MAX_LISTS		32768
+#define IVFFLAT_DEFAULT_PROBES	1
+
+/* Build phases */
+/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
+#define PROGRESS_IVFFLAT_PHASE_KMEANS	2
+#define PROGRESS_IVFFLAT_PHASE_ASSIGN	3
+#define PROGRESS_IVFFLAT_PHASE_LOAD		4
 
 #define IVFFLAT_LIST_SIZE(_dim)	(offsetof(IvfflatListData, center) + VECTOR_SIZE(_dim))
 
 #define IvfflatPageGetOpaque(page)	((IvfflatPageOpaque) PageGetSpecialPointer(page))
 #define IvfflatPageGetMeta(page)	((IvfflatMetaPageData *) PageGetContents(page))
 
-#if PG_VERSION_NUM < 100000
-#define ItemPointerGetBlockNumberNoCheck ItemPointerGetBlockNumber
-#define ItemPointerGetOffsetNumberNoCheck ItemPointerGetOffsetNumber
+#ifdef IVFFLAT_BENCH
+#define IvfflatBench(name, code) \
+	do { \
+		instr_time	start; \
+		instr_time	duration; \
+		INSTR_TIME_SET_CURRENT(start); \
+		(code); \
+		INSTR_TIME_SET_CURRENT(duration); \
+		INSTR_TIME_SUBTRACT(duration, start); \
+		elog(INFO, "%s: %.3f ms", name, INSTR_TIME_GET_MILLISEC(duration)); \
+	} while (0)
+#else
+#define IvfflatBench(name, code) (code)
+#endif
+
+#if PG_VERSION_NUM >= 150000
+#define RandomDouble() pg_prng_double(&pg_global_prng_state)
+#define RandomInt() pg_prng_uint32(&pg_global_prng_state)
+#else
+#define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
+#define RandomInt() random()
 #endif
 
 /* Variables */
-int			ivfflat_probes;
+extern int	ivfflat_probes;
 
 typedef struct VectorArrayData
 {
 	int			length;
 	int			maxlen;
 	int			dim;
-	Vector		items[FLEXIBLE_ARRAY_MEMBER];
+	Vector	   *items;
 }			VectorArrayData;
 
 typedef VectorArrayData * VectorArray;
@@ -63,6 +109,56 @@ typedef struct IvfflatOptions
 	int			lists;			/* number of lists */
 }			IvfflatOptions;
 
+typedef struct IvfflatSpool
+{
+	Tuplesortstate *sortstate;
+	Relation	heap;
+	Relation	index;
+}			IvfflatSpool;
+
+typedef struct IvfflatShared
+{
+	/* Immutable state */
+	Oid			heaprelid;
+	Oid			indexrelid;
+	bool		isconcurrent;
+	int			scantuplesortstates;
+
+	/* Worker progress */
+	ConditionVariable workersdonecv;
+
+	/* Mutex for mutable state */
+	slock_t		mutex;
+
+	/* Mutable state */
+	int			nparticipantsdone;
+	double		reltuples;
+	double		indtuples;
+
+#ifdef IVFFLAT_KMEANS_DEBUG
+	double		inertia;
+#endif
+
+#if PG_VERSION_NUM < 120000
+	ParallelHeapScanDescData heapdesc;	/* must come last */
+#endif
+}			IvfflatShared;
+
+#if PG_VERSION_NUM >= 120000
+#define ParallelTableScanFromIvfflatShared(shared) \
+	(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(IvfflatShared)))
+#endif
+
+typedef struct IvfflatLeader
+{
+	ParallelContext *pcxt;
+	int			nparticipanttuplesorts;
+	IvfflatShared *ivfshared;
+	Sharedsort *sharedsort;
+	Snapshot	snapshot;
+	Vector	   *ivfcenters;
+}			IvfflatLeader;
+
 typedef struct IvfflatBuildState
 {
 	/* Info */
@@ -81,6 +177,7 @@ typedef struct IvfflatBuildState
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *kmeansnormprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -89,6 +186,12 @@ typedef struct IvfflatBuildState
 	ListInfo   *listInfo;
 	Vector	   *normvec;
 
+#ifdef IVFFLAT_KMEANS_DEBUG
+	double		inertia;
+	double	   *listSums;
+	int		   *listCounts;
+#endif
+
 	/* Sampling */
 	BlockSamplerData bs;
 	ReservoirStateData rstate;
@@ -98,6 +201,12 @@ typedef struct IvfflatBuildState
 	Tuplesortstate *sortstate;
 	TupleDesc	tupdesc;
 	TupleTableSlot *slot;
+
+	/* Memory */
+	MemoryContext tmpCtx;
+
+	/* Parallel builds */
+	IvfflatLeader *ivfleader;
 }			IvfflatBuildState;
 
 typedef struct IvfflatMetaPageData
@@ -130,6 +239,7 @@ typedef IvfflatListData * IvfflatList;
 
 typedef struct IvfflatScanList
 {
+	pairingheap_node ph_node;
 	BlockNumber startPage;
 	double		distance;
 }			IvfflatScanList;
@@ -137,8 +247,11 @@ typedef struct IvfflatScanList
 typedef struct IvfflatScanOpaqueData
 {
 	int			probes;
+	int			dimensions;
 	bool		first;
+	bool		hasLock;
 	Buffer		buf;
+	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -151,37 +264,44 @@ typedef struct IvfflatScanOpaqueData
 	FmgrInfo   *normprocinfo;
 	Oid			collation;
 
+	/* Lists */
+	pairingheap *listQueue;
 	IvfflatScanList lists[FLEXIBLE_ARRAY_MEMBER];	/* must come last */
 }			IvfflatScanOpaqueData;
 
 typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 
-#define VECTOR_ARRAY_SIZE(_length, _dim) (offsetof(VectorArrayData, items) + _length * VECTOR_SIZE(_dim))
-#define VECTOR_ARRAY_OFFSET(_arr, _offset) ((char*) _arr + offsetof(VectorArrayData, items) + (_offset) * VECTOR_SIZE(_arr->dim))
+#define VECTOR_ARRAY_SIZE(_length, _dim) (sizeof(VectorArrayData) + (_length) * VECTOR_SIZE(_dim))
+#define VECTOR_ARRAY_OFFSET(_arr, _offset) ((char*) (_arr)->items + (_offset) * VECTOR_SIZE((_arr)->dim))
 #define VectorArrayGet(_arr, _offset) ((Vector *) VECTOR_ARRAY_OFFSET(_arr, _offset))
-#define VectorArraySet(_arr, _offset, _val) (memcpy(VECTOR_ARRAY_OFFSET(_arr, _offset), _val, VECTOR_SIZE(_arr->dim)))
+#define VectorArraySet(_arr, _offset, _val) memcpy(VECTOR_ARRAY_OFFSET(_arr, _offset), _val, VECTOR_SIZE((_arr)->dim))
 
 /* Methods */
-void		_PG_init(void);
 VectorArray VectorArrayInit(int maxlen, int dimensions);
+void		VectorArrayFree(VectorArray arr);
 void		PrintVectorArray(char *msg, VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation rel, uint16 procnum);
 bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 int			IvfflatGetLists(Relation index);
-void		IvfflatUpdateList(Relation index, GenericXLogState *state, ListInfo listInfo, BlockNumber insertPage, BlockNumber startPage, ForkNumber forkNum);
+void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
+void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);
 void		IvfflatCommitBuffer(Buffer buf, GenericXLogState *state);
 void		IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum);
 Buffer		IvfflatNewBuffer(Relation index, ForkNumber forkNum);
-void		IvfflatInitPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
+void		IvfflatInitPage(Buffer buf, Page page);
+void		IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
+void		IvfflatInit(void);
+PGDLLEXPORT void IvfflatParallelBuildMain(dsm_segment *seg, shm_toc *toc);
 
 /* Index access methods */
 IndexBuildResult *ivfflatbuild(Relation heap, Relation index, IndexInfo *indexInfo);
 void		ivfflatbuildempty(Relation index);
 bool		ivfflatinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heap, IndexUniqueCheck checkUnique
-#if PG_VERSION_NUM >= 100000
-						  ,IndexInfo *indexInfo
+#if PG_VERSION_NUM >= 140000
+						  ,bool indexUnchanged
 #endif
+						  ,IndexInfo *indexInfo
 );
 IndexBulkDeleteResult *ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats, IndexBulkDeleteCallback callback, void *callback_state);
 IndexBulkDeleteResult *ivfflatvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats);

src/ivfinsert.c

@@ -4,6 +4,8 @@
 
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "storage/lmgr.h"
+#include "utils/memutils.h"
 
 /*
  * Find the list that minimizes the distance function
@@ -22,6 +24,10 @@ FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo *
 	OffsetNumber offno;
 	OffsetNumber maxoffno;
 
+	/* Avoid compiler warning */
+	listInfo->blkno = nextblkno;
+	listInfo->offno = FirstOffsetNumber;
+
 	procinfo = index_getprocinfo(rel, 1, IVFFLAT_DISTANCE_PROC);
 	collation = rel->rd_indcollation[0];
 
@@ -38,7 +44,7 @@ FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo *
 			list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
 			distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, values[0], PointerGetDatum(&list->center)));
 
-			if (distance < minDistance)
+			if (distance < minDistance || !BlockNumberIsValid(*insertPage))
 			{
 				*insertPage = list->insertPage;
 				listInfo->blkno = nextblkno;
@@ -53,44 +59,59 @@ FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo *
 	}
 }
 
-/*
- * Prepare to insert an index tuple
- */
-static void
-LoadInsertPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, BlockNumber insertPage)
-{
-	*buf = ReadBuffer(index, insertPage);
-	LockBuffer(*buf, BUFFER_LOCK_EXCLUSIVE);
-	*state = GenericXLogStart(index);
-	*page = GenericXLogRegisterBuffer(*state, *buf, 0);
-}
-
 /*
  * Insert a tuple into the index
  */
 static void
-InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
+InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
 {
+	IndexTuple	itup;
+	Datum		value;
+	FmgrInfo   *normprocinfo;
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
 	Size		itemsz;
 	BlockNumber insertPage = InvalidBlockNumber;
 	ListInfo	listInfo;
-	bool		newPage = false;
+	BlockNumber originalInsertPage;
+
+	/* Detoast once for all calls */
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Normalize if needed */
+	normprocinfo = IvfflatOptionalProcInfo(rel, IVFFLAT_NORM_PROC);
+	if (normprocinfo != NULL)
+	{
+		if (!IvfflatNormValue(normprocinfo, rel->rd_indcollation[0], &value, NULL))
+			return;
+	}
 
 	/* Find the insert page - sets the page and list info */
 	FindInsertPage(rel, values, &insertPage, &listInfo);
 	Assert(BlockNumberIsValid(insertPage));
+	originalInsertPage = insertPage;
 
+	/* Form tuple */
+	itup = index_form_tuple(RelationGetDescr(rel), &value, isnull);
+	itup->t_tid = *heap_tid;
+
+	/* Get tuple size */
 	itemsz = MAXALIGN(IndexTupleSize(itup));
 	Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
 
-	LoadInsertPage(rel, &buf, &page, &state, insertPage);
-
 	/* Find a page to insert the item */
-	while (PageGetFreeSpace(page) < itemsz)
+	for (;;)
 	{
+		buf = ReadBuffer(rel, insertPage);
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+
+		state = GenericXLogStart(rel);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+
+		if (PageGetFreeSpace(page) >= itemsz)
+			break;
+
 		insertPage = IvfflatPageGetOpaque(page)->nextblkno;
 
 		if (BlockNumberIsValid(insertPage))
@@ -98,16 +119,40 @@ InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
 			/* Move to next page */
 			GenericXLogAbort(state);
 			UnlockReleaseBuffer(buf);
-
-			LoadInsertPage(rel, &buf, &page, &state, insertPage);
 		}
 		else
 		{
-			/* Add a new page */
-			IvfflatAppendPage(rel, &buf, &page, &state, MAIN_FORKNUM);
+			Buffer		newbuf;
+			Page		newpage;
 
-			insertPage = BufferGetBlockNumber(buf);
-			newPage = true;
+			/* Add a new page */
+			LockRelationForExtension(rel, ExclusiveLock);
+			newbuf = IvfflatNewBuffer(rel, MAIN_FORKNUM);
+			UnlockRelationForExtension(rel, ExclusiveLock);
+
+			/* Init new page */
+			newpage = GenericXLogRegisterBuffer(state, newbuf, GENERIC_XLOG_FULL_IMAGE);
+			IvfflatInitPage(newbuf, newpage);
+
+			/* Update insert page */
+			insertPage = BufferGetBlockNumber(newbuf);
+
+			/* Update previous buffer */
+			IvfflatPageGetOpaque(page)->nextblkno = insertPage;
+
+			/* Commit */
+			MarkBufferDirty(newbuf);
+			MarkBufferDirty(buf);
+			GenericXLogFinish(state);
+
+			/* Unlock previous buffer */
+			UnlockReleaseBuffer(buf);
+
+			/* Prepare new buffer */
+			state = GenericXLogStart(rel);
+			buf = newbuf;
+			page = GenericXLogRegisterBuffer(state, buf, 0);
+			break;
 		}
 	}
 
@@ -118,8 +163,8 @@ InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
 	IvfflatCommitBuffer(buf, state);
 
 	/* Update the insert page */
-	if (newPage)
-		IvfflatUpdateList(rel, state, listInfo, insertPage, InvalidBlockNumber, MAIN_FORKNUM);
+	if (insertPage != originalInsertPage)
+		IvfflatUpdateList(rel, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
 }
 
 /*
@@ -128,36 +173,34 @@ InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
 bool
 ivfflatinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 			  Relation heap, IndexUniqueCheck checkUnique
-#if PG_VERSION_NUM >= 100000
-			  ,IndexInfo *indexInfo
+#if PG_VERSION_NUM >= 140000
+			  ,bool indexUnchanged
 #endif
+			  ,IndexInfo *indexInfo
 )
 {
-	IndexTuple	itup;
-	Datum		value;
-	FmgrInfo   *normprocinfo;
+	MemoryContext oldCtx;
+	MemoryContext insertCtx;
 
+	/* Skip nulls */
 	if (isnull[0])
 		return false;
 
-	value = values[0];
+	/*
+	 * Use memory context since detoast, IvfflatNormValue, and
+	 * index_form_tuple can allocate
+	 */
+	insertCtx = AllocSetContextCreate(CurrentMemoryContext,
+									  "Ivfflat insert temporary context",
+									  ALLOCSET_DEFAULT_SIZES);
+	oldCtx = MemoryContextSwitchTo(insertCtx);
 
-	/* Normalize if needed */
-	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
-	if (normprocinfo != NULL)
-	{
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
-			return false;
-	}
+	/* Insert tuple */
+	InsertTuple(index, values, isnull, heap_tid, heap);
 
-	itup = index_form_tuple(RelationGetDescr(index), &value, isnull);
-	itup->t_tid = *heap_tid;
-	InsertTuple(index, itup, heap, &value);
-	pfree(itup);
-
-	/* Clean up if we allocated a new value */
-	if (value != values[0])
-		pfree(DatumGetPointer(value));
+	/* Delete memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextDelete(insertCtx);
 
 	return false;
 }

src/ivfkmeans.c

@@ -1,6 +1,7 @@
 #include "postgres.h"
 
 #include <float.h>
+#include <math.h>
 
 #include "ivfflat.h"
 #include "miscadmin.h"
@@ -11,17 +12,16 @@
  * https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf
  */
 static void
-InitCenters(Relation index, VectorArray samples, VectorArray centers, double *lowerBound)
+InitCenters(Relation index, VectorArray samples, VectorArray centers, float *lowerBound)
 {
 	FmgrInfo   *procinfo;
 	Oid			collation;
-	int			i;
-	int			j;
+	int64		j;
 	double		distance;
 	double		sum;
 	double		choice;
 	Vector	   *vec;
-	double	   *weight = palloc(samples->length * sizeof(double));
+	float	   *weight = palloc(samples->length * sizeof(float));
 	int			numCenters = centers->maxlen;
 	int			numSamples = samples->length;
 
@@ -29,13 +29,13 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, double *lo
 	collation = index->rd_indcollation[0];
 
 	/* Choose an initial center uniformly at random */
-	VectorArraySet(centers, 0, VectorArrayGet(samples, random() % samples->length));
+	VectorArraySet(centers, 0, VectorArrayGet(samples, RandomInt() % samples->length));
 	centers->length++;
 
 	for (j = 0; j < numSamples; j++)
 		weight[j] = DBL_MAX;
 
-	for (i = 0; i < numCenters; i++)
+	for (int i = 0; i < numCenters; i++)
 	{
 		CHECK_FOR_INTERRUPTS();
 
@@ -66,7 +66,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, double *lo
 			break;
 
 		/* Choose new center using weighted probability distribution. */
-		choice = sum * (((double) random()) / MAX_RANDOM_VALUE);
+		choice = sum * RandomDouble();
 		for (j = 0; j < numSamples - 1; j++)
 		{
 			choice -= weight[j];
@@ -87,13 +87,12 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, double *lo
 static inline void
 ApplyNorm(FmgrInfo *normprocinfo, Oid collation, Vector * vec)
 {
-	int			i;
 	double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(vec)));
 
 	/* TODO Handle zero norm */
 	if (norm > 0)
 	{
-		for (i = 0; i < vec->dim; i++)
+		for (int i = 0; i < vec->dim; i++)
 			vec->x[i] /= norm;
 	}
 }
@@ -113,23 +112,24 @@ CompareVectors(const void *a, const void *b)
 static void
 QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 {
-	int			i;
-	int			j;
 	Vector	   *vec;
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 
 	/* Copy existing vectors while avoiding duplicates */
-	qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
-	for (i = 0; i < samples->length; i++)
+	if (samples->length > 0)
 	{
-		vec = VectorArrayGet(samples, i);
-
-		if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
+		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
+		for (int i = 0; i < samples->length; i++)
 		{
-			VectorArraySet(centers, centers->length, vec);
-			centers->length++;
+			vec = VectorArrayGet(samples, i);
+
+			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
+			{
+				VectorArraySet(centers, centers->length, vec);
+				centers->length++;
+			}
 		}
 	}
 
@@ -141,8 +141,8 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 		SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
 		vec->dim = dimensions;
 
-		for (j = 0; j < dimensions; j++)
-			vec->x[j] = ((double) random()) / MAX_RANDOM_VALUE;
+		for (int j = 0; j < dimensions; j++)
+			vec->x[j] = RandomDouble();
 
 		/* Normalize if needed (only needed for random centers) */
 		if (normprocinfo != NULL)
@@ -169,19 +169,19 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	Vector	   *vec;
 	Vector	   *newCenter;
 	int			iteration;
-	int			j;
-	int			k;
+	int64		j;
+	int64		k;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
 	int			numSamples = samples->length;
 	VectorArray newCenters;
 	int		   *centerCounts;
 	int		   *closestCenters;
-	double	   *lowerBound;
-	double	   *upperBound;
-	double	   *s;
-	double	   *halfcdist;
-	double	   *newcdist;
+	float	   *lowerBound;
+	float	   *upperBound;
+	float	   *s;
+	float	   *halfcdist;
+	float	   *newcdist;
 	int			changes;
 	double		minDistance;
 	int			closestCenter;
@@ -191,19 +191,47 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	double		dxcx;
 	double		dxc;
 
+	/* Calculate allocation sizes */
+	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
+	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->dim);
+	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, dimensions);
+	Size		centerCountsSize = sizeof(int) * numCenters;
+	Size		closestCentersSize = sizeof(int) * numSamples;
+	Size		lowerBoundSize = sizeof(float) * numSamples * numCenters;
+	Size		upperBoundSize = sizeof(float) * numSamples;
+	Size		sSize = sizeof(float) * numCenters;
+	Size		halfcdistSize = sizeof(float) * numCenters * numCenters;
+	Size		newcdistSize = sizeof(float) * numCenters;
+
+	/* Calculate total size */
+	Size		totalSize = samplesSize + centersSize + newCentersSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
+
+	/* Check memory requirements */
+	/* Add one to error message to ceil */
+	if (totalSize > (Size) maintenance_work_mem * 1024L)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("memory required is %zu MB, maintenance_work_mem is %d MB",
+						totalSize / (1024 * 1024) + 1, maintenance_work_mem / 1024)));
+
+	/* Ensure indexing does not overflow */
+	if (numCenters * numCenters > INT_MAX)
+		elog(ERROR, "Indexing overflow detected. Please report a bug.");
+
 	/* Set support functions */
 	procinfo = index_getprocinfo(index, 1, IVFFLAT_KMEANS_DISTANCE_PROC);
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 	collation = index->rd_indcollation[0];
 
 	/* Allocate space */
-	centerCounts = palloc(sizeof(int) * numCenters);
-	closestCenters = palloc(sizeof(int) * numSamples);
-	lowerBound = palloc(sizeof(double) * numSamples * numCenters);
-	upperBound = palloc(sizeof(double) * numSamples);
-	s = palloc(sizeof(double) * numCenters);
-	halfcdist = palloc(sizeof(double) * numCenters * numCenters);
-	newcdist = palloc(sizeof(double) * numCenters);
+	/* Use float instead of double to save memory */
+	centerCounts = palloc(centerCountsSize);
+	closestCenters = palloc(closestCentersSize);
+	lowerBound = palloc_extended(lowerBoundSize, MCXT_ALLOC_HUGE);
+	upperBound = palloc(upperBoundSize);
+	s = palloc(sSize);
+	halfcdist = palloc_extended(halfcdistSize, MCXT_ALLOC_HUGE);
+	newcdist = palloc(newcdistSize);
 
 	newCenters = VectorArrayInit(numCenters, dimensions);
 	for (j = 0; j < numCenters; j++)
@@ -220,9 +248,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	for (j = 0; j < numSamples; j++)
 	{
 		minDistance = DBL_MAX;
-		closestCenter = -1;
-
-		vec = VectorArrayGet(samples, j);
+		closestCenter = 0;
 
 		/* Find closest center */
 		for (k = 0; k < numCenters; k++)
@@ -369,6 +395,14 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 			if (centerCounts[j] > 0)
 			{
+				/* Double avoids overflow, but requires more memory */
+				/* TODO Update bounds */
+				for (k = 0; k < dimensions; k++)
+				{
+					if (isinf(vec->x[k]))
+						vec->x[k] = vec->x[k] > 0 ? FLT_MAX : -FLT_MAX;
+				}
+
 				for (k = 0; k < dimensions; k++)
 					vec->x[k] /= centerCounts[j];
 			}
@@ -376,7 +410,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 			{
 				/* TODO Handle empty centers properly */
 				for (k = 0; k < dimensions; k++)
-					vec->x[k] = ((double) random()) / MAX_RANDOM_VALUE;
+					vec->x[k] = RandomDouble();
 			}
 
 			/* Normalize if needed */
@@ -414,7 +448,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 			break;
 	}
 
-	pfree(newCenters);
+	VectorArrayFree(newCenters);
 	pfree(centerCounts);
 	pfree(closestCenters);
 	pfree(lowerBound);
@@ -432,16 +466,31 @@ CheckCenters(Relation index, VectorArray centers)
 {
 	FmgrInfo   *normprocinfo;
 	Oid			collation;
-	int			i;
+	Vector	   *vec;
 	double		norm;
 
 	if (centers->length != centers->maxlen)
 		elog(ERROR, "Not enough centers. Please report a bug.");
 
+	/* Ensure no NaN or infinite values */
+	for (int i = 0; i < centers->length; i++)
+	{
+		vec = VectorArrayGet(centers, i);
+
+		for (int j = 0; j < vec->dim; j++)
+		{
+			if (isnan(vec->x[j]))
+				elog(ERROR, "NaN detected. Please report a bug.");
+
+			if (isinf(vec->x[j]))
+				elog(ERROR, "Infinite value detected. Please report a bug.");
+		}
+	}
+
 	/* Ensure no duplicate centers */
 	/* Fine to sort in-place */
 	qsort(centers->items, centers->length, VECTOR_SIZE(centers->dim), CompareVectors);
-	for (i = 1; i < centers->length; i++)
+	for (int i = 1; i < centers->length; i++)
 	{
 		if (CompareVectors(VectorArrayGet(centers, i), VectorArrayGet(centers, i - 1)) == 0)
 			elog(ERROR, "Duplicate centers detected. Please report a bug.");
@@ -454,7 +503,7 @@ CheckCenters(Relation index, VectorArray centers)
 	{
 		collation = index->rd_indcollation[0];
 
-		for (i = 0; i < centers->length; i++)
+		for (int i = 0; i < centers->length; i++)
 		{
 			norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
 			if (norm == 0)

src/ivfscan.c

@@ -1,30 +1,26 @@
 #include "postgres.h"
 
-#include "access/relscan.h"
-#include "ivfflat.h"
-#include "miscadmin.h"
-#include "storage/bufmgr.h"
+#include <float.h>
 
-#if PG_VERSION_NUM >= 110000
+#include "access/relscan.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
-#else
-#include "catalog/pg_operator.h"
-#include "catalog/pg_type.h"
-#endif
+#include "ivfflat.h"
+#include "miscadmin.h"
+#include "pgstat.h"
+#include "storage/bufmgr.h"
+#include "storage/lmgr.h"
 
 /*
  * Compare list distances
  */
 static int
-CompareLists(const void *a, const void *b)
+CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
 {
-	double		diff = (((IvfflatScanList *) a)->distance - ((IvfflatScanList *) b)->distance);
-
-	if (diff > 0)
+	if (((const IvfflatScanList *) a)->distance > ((const IvfflatScanList *) b)->distance)
 		return 1;
 
-	if (diff < 0)
+	if (((const IvfflatScanList *) a)->distance < ((const IvfflatScanList *) b)->distance)
 		return -1;
 
 	return 0;
@@ -36,47 +32,69 @@ CompareLists(const void *a, const void *b)
 static void
 GetScanLists(IndexScanDesc scan, Datum value)
 {
-	Buffer		cbuf;
-	Page		cpage;
-	IvfflatList list;
-	OffsetNumber offno;
-	OffsetNumber maxoffno;
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 	BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
 	int			listCount = 0;
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	double		distance;
+	double		maxDistance = DBL_MAX;
 
 	/* Search all list pages */
 	while (BlockNumberIsValid(nextblkno))
 	{
+		Buffer		cbuf;
+		Page		cpage;
+		OffsetNumber maxoffno;
+
 		cbuf = ReadBuffer(scan->indexRelation, nextblkno);
 		LockBuffer(cbuf, BUFFER_LOCK_SHARE);
 		cpage = BufferGetPage(cbuf);
 
 		maxoffno = PageGetMaxOffsetNumber(cpage);
 
-		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
+			IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
+			double		distance;
 
 			/* Use procinfo from the index instead of scan key for performance */
 			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
-			so->lists[listCount].startPage = list->startPage;
-			so->lists[listCount].distance = distance;
-			listCount++;
+			if (listCount < so->probes)
+			{
+				IvfflatScanList *scanlist;
+
+				scanlist = &so->lists[listCount];
+				scanlist->startPage = list->startPage;
+				scanlist->distance = distance;
+				listCount++;
+
+				/* Add to heap */
+				pairingheap_add(so->listQueue, &scanlist->ph_node);
+
+				/* Calculate max distance */
+				if (listCount == so->probes)
+					maxDistance = ((IvfflatScanList *) pairingheap_first(so->listQueue))->distance;
+			}
+			else if (distance < maxDistance)
+			{
+				IvfflatScanList *scanlist;
+
+				/* Remove */
+				scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
+
+				/* Reuse */
+				scanlist->startPage = list->startPage;
+				scanlist->distance = distance;
+				pairingheap_add(so->listQueue, &scanlist->ph_node);
+
+				/* Update max distance */
+				maxDistance = ((IvfflatScanList *) pairingheap_first(so->listQueue))->distance;
+			}
 		}
 
 		nextblkno = IvfflatPageGetOpaque(cpage)->nextblkno;
 
 		UnlockReleaseBuffer(cbuf);
 	}
-
-	/* Sort by distance */
-	qsort(so->lists, listCount, sizeof(IvfflatScanList), CompareLists);
-
-	if (so->probes > listCount)
-		so->probes = listCount;
 }
 
 /*
@@ -86,16 +104,8 @@ static void
 GetScanItems(IndexScanDesc scan, Datum value)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	Buffer		buf;
-	Page		page;
-	IndexTuple	itup;
-	BlockNumber searchPage;
-	OffsetNumber offno;
-	OffsetNumber maxoffno;
-	Datum		datum;
-	bool		isnull;
-	int			i;
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
+	double		tuples = 0;
 
 #if PG_VERSION_NUM >= 120000
 	TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
@@ -111,21 +121,34 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	BufferAccessStrategy bas = GetAccessStrategy(BAS_BULKREAD);
 
 	/* Search closest probes lists */
-	for (i = 0; i < so->probes; i++)
+	while (!pairingheap_is_empty(so->listQueue))
 	{
-		searchPage = so->lists[i].startPage;
+		BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
 
 		/* Search all entry pages for list */
 		while (BlockNumberIsValid(searchPage))
 		{
+			Buffer		buf;
+			Page		page;
+			OffsetNumber maxoffno;
+
 			buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
 			LockBuffer(buf, BUFFER_LOCK_SHARE);
 			page = BufferGetPage(buf);
 			maxoffno = PageGetMaxOffsetNumber(page);
 
-			for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+			for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 			{
-				itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
+				IndexTuple	itup;
+				Datum		datum;
+				bool		isnull;
+				ItemId		itemid = PageGetItemId(page, offno);
+
+				/* Skip dead tuples */
+				if (scan->ignore_killed_tuples && ItemIdIsDead(itemid))
+					continue;
+
+				itup = (IndexTuple) PageGetItem(page, itemid);
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
 
 				/*
@@ -137,15 +160,15 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				ExecClearTuple(slot);
 				slot->tts_values[0] = FunctionCall2Coll(so->procinfo, so->collation, datum, value);
 				slot->tts_isnull[0] = false;
-				slot->tts_values[1] = Int32GetDatum((int) ItemPointerGetBlockNumberNoCheck(&itup->t_tid));
+				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
-				slot->tts_values[2] = Int32GetDatum((int) ItemPointerGetOffsetNumberNoCheck(&itup->t_tid));
+				slot->tts_values[2] = Int32GetDatum((int) searchPage);
 				slot->tts_isnull[2] = false;
-				slot->tts_values[3] = Int32GetDatum((int) searchPage);
-				slot->tts_isnull[3] = false;
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
+
+				tuples++;
 			}
 
 			searchPage = IvfflatPageGetOpaque(page)->nextblkno;
@@ -153,6 +176,65 @@ GetScanItems(IndexScanDesc scan, Datum value)
 			UnlockReleaseBuffer(buf);
 		}
 	}
+
+	FreeAccessStrategy(bas);
+
+	if (tuples < 100)
+		ereport(DEBUG1,
+				(errmsg("index scan found few tuples"),
+				 errdetail("Index may have been created with little data."),
+				 errhint("Recreate the index and possibly decrease lists.")));
+
+	tuplesort_performsort(so->sortstate);
+}
+
+/*
+ * Mark prior tuple as dead
+ */
+static void
+MarkPriorTupleDead(IndexScanDesc scan)
+{
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	Buffer		buf = so->buf;
+	Page		page;
+	OffsetNumber maxoffno;
+
+	/* Safety check */
+	if (!BufferIsValid(so->buf) || !ItemPointerIsValid(&so->heaptid))
+		return;
+
+	/* Only a shared locked is needed for ItemIdMarkDead */
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	maxoffno = PageGetMaxOffsetNumber(page);
+
+	for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+	{
+		ItemId		itemid = PageGetItemId(page, offno);
+		IndexTuple	itup = (IndexTuple) PageGetItem(page, itemid);
+
+		/*
+		 * Find tuple. Since buffer has been pinned, tuple cannot have been
+		 * vacuumed (and heap TID reused).
+		 */
+		if (ItemPointerEquals(&itup->t_tid, &so->heaptid))
+		{
+			/*
+			 * Make sure tuple has not already been marked dead to avoid extra
+			 * WAL if wal_log_hints or data checksums enabled
+			 */
+			if (!ItemIdIsDead(itemid))
+			{
+				ItemIdMarkDead(itemid);
+				MarkBufferDirtyHint(buf, true);
+			}
+
+			break;
+		}
+	}
+
+	/* Unlock buffer */
+	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
 }
 
 /*
@@ -164,17 +246,28 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	IndexScanDesc scan;
 	IvfflatScanOpaque so;
 	int			lists;
+	int			dimensions;
 	AttrNumber	attNums[] = {1};
 	Oid			sortOperators[] = {Float8LessOperator};
 	Oid			sortCollations[] = {InvalidOid};
 	bool		nullsFirstFlags[] = {false};
+	int			probes = ivfflat_probes;
 
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
-	lists = IvfflatGetLists(scan->indexRelation);
 
-	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + lists * sizeof(IvfflatScanList));
+	/* Get lists and dimensions from metapage */
+	IvfflatGetMetaPageInfo(index, &lists, &dimensions);
+
+	if (probes > lists)
+		probes = lists;
+
+	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
 	so->buf = InvalidBuffer;
 	so->first = true;
+	so->hasLock = false;
+	ItemPointerSetInvalid(&so->heaptid);
+	so->probes = probes;
+	so->dimensions = dimensions;
 
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -183,21 +276,16 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(4);
+	so->tupdesc = CreateTemplateTupleDesc(3);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(4, false);
+	so->tupdesc = CreateTemplateTupleDesc(3, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "blkno", INT4OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "offset", INT4OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 4, "indexblkno", INT4OID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
-#if PG_VERSION_NUM >= 110000
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
-#else
-	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, false);
-#endif
 
 #if PG_VERSION_NUM >= 120000
 	so->slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsMinimalTuple);
@@ -205,6 +293,8 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->slot = MakeSingleTupleTableSlot(so->tupdesc);
 #endif
 
+	so->listQueue = pairingheap_allocate(CompareLists, scan);
+
 	scan->opaque = so;
 
 	return scan;
@@ -224,7 +314,8 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 #endif
 
 	so->first = true;
-	so->probes = ivfflat_probes;
+	ItemPointerSetInvalid(&so->heaptid);
+	pairingheap_reset(so->listQueue);
 
 	if (keys && scan->numberOfKeys > 0)
 		memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));
@@ -251,45 +342,65 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 	{
 		Datum		value;
 
-		/* No items will match if null */
-		if (scan->orderByData->sk_flags & SK_ISNULL)
-			return false;
+		/* Count index scan for stats */
+		pgstat_count_index_scan(scan->indexRelation);
 
-		value = scan->orderByData->sk_argument;
+		/* Safety check */
+		if (scan->orderByData == NULL)
+			elog(ERROR, "cannot scan ivfflat index without order");
 
-		if (so->normprocinfo != NULL)
+		/* Get a shared lock for non-MVCC snapshots */
+		if (!so->hasLock && !IsMVCCSnapshot(scan->xs_snapshot))
 		{
-			/* No items will match if normalization fails */
-			if (!IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL))
-				return false;
+			so->hasLock = true;
+			LockPage(scan->indexRelation, IVFFLAT_SCAN_LOCK, ShareLock);
 		}
 
-		GetScanLists(scan, value);
-		GetScanItems(scan, value);
-		tuplesort_performsort(so->sortstate);
+		if (scan->orderByData->sk_flags & SK_ISNULL)
+			value = PointerGetDatum(InitVector(so->dimensions));
+		else
+		{
+			value = scan->orderByData->sk_argument;
+
+			/* Value should not be compressed or toasted */
+			Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+			Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
+
+			/* Fine if normalization fails */
+			if (so->normprocinfo != NULL)
+				IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL);
+		}
+
+		IvfflatBench("GetScanLists", GetScanLists(scan, value));
+		IvfflatBench("GetScanItems", GetScanItems(scan, value));
 		so->first = false;
 
 		/* Clean up if we allocated a new value */
 		if (value != scan->orderByData->sk_argument)
 			pfree(DatumGetPointer(value));
 	}
-
-#if PG_VERSION_NUM >= 100000
-	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
-#else
-	if (tuplesort_gettupleslot(so->sortstate, true, so->slot, NULL))
-#endif
+	else
 	{
-		BlockNumber blkno = DatumGetInt32(slot_getattr(so->slot, 2, &so->isnull));
-		OffsetNumber offset = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
-		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 4, &so->isnull));
+		/* Mark prior tuple as dead */
+		if (scan->kill_prior_tuple)
+			MarkPriorTupleDead(scan);
+	}
+
+	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
+	{
+		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
+		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
 
 #if PG_VERSION_NUM >= 120000
-		ItemPointerSet(&scan->xs_heaptid, blkno, offset);
+		scan->xs_heaptid = *heaptid;
 #else
-		ItemPointerSet(&scan->xs_ctup.t_self, blkno, offset);
+		scan->xs_ctup.t_self = *heaptid;
 #endif
 
+		/* Keep track of info needed to mark tuple as dead */
+		so->heaptid = *heaptid;
+
+		/* Unpin buffer */
 		if (BufferIsValid(so->buf))
 			ReleaseBuffer(so->buf);
 
@@ -320,6 +431,11 @@ ivfflatendscan(IndexScanDesc scan)
 	if (BufferIsValid(so->buf))
 		ReleaseBuffer(so->buf);
 
+	/* Release lock */
+	if (so->hasLock)
+		UnlockPage(scan->indexRelation, IVFFLAT_SCAN_LOCK, ShareLock);
+
+	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 
 	pfree(so);

src/ivfutils.c

@@ -10,23 +10,32 @@
 VectorArray
 VectorArrayInit(int maxlen, int dimensions)
 {
-	VectorArray res = palloc0(VECTOR_ARRAY_SIZE(maxlen, dimensions));
+	VectorArray res = palloc(sizeof(VectorArrayData));
 
 	res->length = 0;
 	res->maxlen = maxlen;
 	res->dim = dimensions;
+	res->items = palloc_extended(maxlen * VECTOR_SIZE(dimensions), MCXT_ALLOC_ZERO | MCXT_ALLOC_HUGE);
 	return res;
 }
 
+/*
+ * Free a vector array
+ */
+void
+VectorArrayFree(VectorArray arr)
+{
+	pfree(arr->items);
+	pfree(arr);
+}
+
 /*
  * Print vector array - useful for debugging
  */
 void
 PrintVectorArray(char *msg, VectorArray arr)
 {
-	int			i;
-
-	for (i = 0; i < arr->length; i++)
+	for (int i = 0; i < arr->length; i++)
 		PrintVector(msg, VectorArrayGet(arr, i));
 }
 
@@ -67,20 +76,16 @@ IvfflatOptionalProcInfo(Relation rel, uint16 procnum)
 bool
 IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 {
-	Vector	   *v;
-	int			i;
-	double		norm;
-
-	norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
+	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
 	{
-		v = (Vector *) DatumGetPointer(*value);
+		Vector	   *v = DatumGetVector(*value);
 
 		if (result == NULL)
 			result = InitVector(v->dim);
 
-		for (i = 0; i < v->dim; i++)
+		for (int i = 0; i < v->dim; i++)
 			result->x[i] = v->x[i] / norm;
 
 		*value = PointerGetDatum(result);
@@ -107,13 +112,22 @@ IvfflatNewBuffer(Relation index, ForkNumber forkNum)
  * Init page
  */
 void
-IvfflatInitPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state)
+IvfflatInitPage(Buffer buf, Page page)
+{
+	PageInit(page, BufferGetPageSize(buf), sizeof(IvfflatPageOpaqueData));
+	IvfflatPageGetOpaque(page)->nextblkno = InvalidBlockNumber;
+	IvfflatPageGetOpaque(page)->page_id = IVFFLAT_PAGE_ID;
+}
+
+/*
+ * Init and register page
+ */
+void
+IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state)
 {
 	*state = GenericXLogStart(index);
 	*page = GenericXLogRegisterBuffer(*state, *buf, GENERIC_XLOG_FULL_IMAGE);
-	PageInit(*page, BufferGetPageSize(*buf), sizeof(IvfflatPageOpaqueData));
-	IvfflatPageGetOpaque(*page)->nextblkno = InvalidBlockNumber;
-	IvfflatPageGetOpaque(*page)->page_id = IVFFLAT_PAGE_ID;
+	IvfflatInitPage(*buf, *page);
 }
 
 /*
@@ -135,29 +149,65 @@ IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
 void
 IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum)
 {
-	Buffer		prevbuf = *buf;
-
 	/* Get new buffer */
-	*buf = IvfflatNewBuffer(index, forkNum);
+	Buffer		newbuf = IvfflatNewBuffer(index, forkNum);
+	Page		newpage = GenericXLogRegisterBuffer(*state, newbuf, GENERIC_XLOG_FULL_IMAGE);
 
-	/* Update and commit previous buffer */
-	IvfflatPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(*buf);
-	IvfflatCommitBuffer(prevbuf, *state);
+	/* Update the previous buffer */
+	IvfflatPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
 
 	/* Init new page */
-	IvfflatInitPage(index, buf, page, state);
+	IvfflatInitPage(newbuf, newpage);
+
+	/* Commit */
+	MarkBufferDirty(*buf);
+	MarkBufferDirty(newbuf);
+	GenericXLogFinish(*state);
+
+	/* Unlock */
+	UnlockReleaseBuffer(*buf);
+
+	*state = GenericXLogStart(index);
+	*page = GenericXLogRegisterBuffer(*state, newbuf, GENERIC_XLOG_FULL_IMAGE);
+	*buf = newbuf;
+}
+
+/*
+ * Get the metapage info
+ */
+void
+IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions)
+{
+	Buffer		buf;
+	Page		page;
+	IvfflatMetaPage metap;
+
+	buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	metap = IvfflatPageGetMeta(page);
+
+	*lists = metap->lists;
+
+	if (dimensions != NULL)
+		*dimensions = metap->dimensions;
+
+	UnlockReleaseBuffer(buf);
 }
 
 /*
  * Update the start or insert page of a list
  */
 void
-IvfflatUpdateList(Relation index, GenericXLogState *state, ListInfo listInfo,
-				  BlockNumber insertPage, BlockNumber startPage, ForkNumber forkNum)
+IvfflatUpdateList(Relation index, ListInfo listInfo,
+				  BlockNumber insertPage, BlockNumber originalInsertPage,
+				  BlockNumber startPage, ForkNumber forkNum)
 {
 	Buffer		buf;
 	Page		page;
+	GenericXLogState *state;
 	IvfflatList list;
+	bool		changed = false;
 
 	buf = ReadBufferExtended(index, forkNum, listInfo.blkno, RBM_NORMAL, NULL);
 	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
@@ -165,12 +215,29 @@ IvfflatUpdateList(Relation index, GenericXLogState *state, ListInfo listInfo,
 	page = GenericXLogRegisterBuffer(state, buf, 0);
 	list = (IvfflatList) PageGetItem(page, PageGetItemId(page, listInfo.offno));
 
-	if (BlockNumberIsValid(insertPage))
-		list->insertPage = insertPage;
+	if (BlockNumberIsValid(insertPage) && insertPage != list->insertPage)
+	{
+		/* Skip update if insert page is lower than original insert page  */
+		/* This is needed to prevent insert from overwriting vacuum */
+		if (!BlockNumberIsValid(originalInsertPage) || insertPage >= originalInsertPage)
+		{
+			list->insertPage = insertPage;
+			changed = true;
+		}
+	}
 
-	if (BlockNumberIsValid(startPage))
+	if (BlockNumberIsValid(startPage) && startPage != list->startPage)
+	{
 		list->startPage = startPage;
+		changed = true;
+	}
 
-	/* Could only commit if changed, but extra complexity isn't needed */
-	IvfflatCommitBuffer(buf, state);
+	/* Only commit if changed */
+	if (changed)
+		IvfflatCommitBuffer(buf, state);
+	else
+	{
+		GenericXLogAbort(state);
+		UnlockReleaseBuffer(buf);
+	}
 }

src/ivfvacuum.c

@@ -3,6 +3,7 @@
 #include "commands/vacuum.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "storage/lmgr.h"
 
 /*
  * Bulk delete tuples from the index
@@ -12,34 +13,23 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 				  IndexBulkDeleteCallback callback, void *callback_state)
 {
 	Relation	index = info->index;
-	Buffer		cbuf;
-	Page		cpage;
-	Buffer		buf;
-	Page		page;
-	IvfflatList list;
-	IndexTuple	itup;
-	ItemPointer htup;
-	OffsetNumber deletable[MaxOffsetNumber];
-	int			ndeletable;
-	OffsetNumber startPages[MaxOffsetNumber];
-	BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
-	BlockNumber searchPage;
-	BlockNumber insertPage;
-	GenericXLogState *state;
-	OffsetNumber coffno;
-	OffsetNumber cmaxoffno;
-	OffsetNumber offno;
-	OffsetNumber maxoffno;
-	ListInfo	listInfo;
+	BlockNumber blkno = IVFFLAT_HEAD_BLKNO;
 	BufferAccessStrategy bas = GetAccessStrategy(BAS_BULKREAD);
 
 	if (stats == NULL)
 		stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
 
 	/* Iterate over list pages */
-	while (BlockNumberIsValid(nextblkno))
+	while (BlockNumberIsValid(blkno))
 	{
-		cbuf = ReadBuffer(index, nextblkno);
+		Buffer		cbuf;
+		Page		cpage;
+		OffsetNumber coffno;
+		OffsetNumber cmaxoffno;
+		BlockNumber startPages[MaxOffsetNumber];
+		ListInfo	listInfo;
+
+		cbuf = ReadBuffer(index, blkno);
 		LockBuffer(cbuf, BUFFER_LOCK_SHARE);
 		cpage = BufferGetPage(cbuf);
 
@@ -48,33 +38,38 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 		/* Iterate over lists */
 		for (coffno = FirstOffsetNumber; coffno <= cmaxoffno; coffno = OffsetNumberNext(coffno))
 		{
-			list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, coffno));
+			IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, coffno));
+
 			startPages[coffno - FirstOffsetNumber] = list->startPage;
 		}
 
-		listInfo.blkno = nextblkno;
-		nextblkno = IvfflatPageGetOpaque(cpage)->nextblkno;
+		listInfo.blkno = blkno;
+		blkno = IvfflatPageGetOpaque(cpage)->nextblkno;
 
 		UnlockReleaseBuffer(cbuf);
 
 		for (coffno = FirstOffsetNumber; coffno <= cmaxoffno; coffno = OffsetNumberNext(coffno))
 		{
-			searchPage = startPages[coffno - FirstOffsetNumber];
-			insertPage = InvalidBlockNumber;
+			BlockNumber searchPage = startPages[coffno - FirstOffsetNumber];
+			BlockNumber insertPage = InvalidBlockNumber;
 
 			/* Iterate over entry pages */
 			while (BlockNumberIsValid(searchPage))
 			{
+				Buffer		buf;
+				Page		page;
+				GenericXLogState *state;
+				OffsetNumber offno;
+				OffsetNumber maxoffno;
+				OffsetNumber deletable[MaxOffsetNumber];
+				int			ndeletable;
+
 				vacuum_delay_point();
 
-				buf = ReadBufferExtended(index, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
+				/* Ensure no in-flight index scans for non-MVCC snapshots */
+				LockPage(index, IVFFLAT_SCAN_LOCK, ExclusiveLock);
 
-				/*
-				 * ambulkdelete cannot delete entries from pages that are
-				 * pinned by other backends
-				 *
-				 * https://www.postgresql.org/docs/current/index-locking.html
-				 */
+				buf = ReadBufferExtended(index, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
 				LockBufferForCleanup(buf);
 
 				state = GenericXLogStart(index);
@@ -86,8 +81,8 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 				/* Find deleted tuples */
 				for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 				{
-					itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
-					htup = &(itup->t_tid);
+					IndexTuple	itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
+					ItemPointer htup = &(itup->t_tid);
 
 					if (callback(htup, callback_state))
 					{
@@ -98,6 +93,11 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 						stats->num_index_tuples++;
 				}
 
+				/* Set to first free page */
+				/* Must be set before searchPage is updated */
+				if (!BlockNumberIsValid(insertPage) && ndeletable > 0)
+					insertPage = searchPage;
+
 				searchPage = IvfflatPageGetOpaque(page)->nextblkno;
 
 				if (ndeletable > 0)
@@ -106,15 +106,13 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 					PageIndexMultiDelete(page, deletable, ndeletable);
 					MarkBufferDirty(buf);
 					GenericXLogFinish(state);
-
-					/* Set to first free page */
-					if (!BlockNumberIsValid(insertPage))
-						insertPage = searchPage;
 				}
 				else
 					GenericXLogAbort(state);
 
 				UnlockReleaseBuffer(buf);
+
+				UnlockPage(index, IVFFLAT_SCAN_LOCK, ExclusiveLock);
 			}
 
 			/*
@@ -123,14 +121,16 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 			 * We don't add or delete items from lists pages, so offset won't
 			 * change.
 			 */
-			if (!BlockNumberIsValid(insertPage))
+			if (BlockNumberIsValid(insertPage))
 			{
 				listInfo.offno = coffno;
-				IvfflatUpdateList(index, state, listInfo, insertPage, InvalidBlockNumber, MAIN_FORKNUM);
+				IvfflatUpdateList(index, listInfo, insertPage, InvalidBlockNumber, InvalidBlockNumber, MAIN_FORKNUM);
 			}
 		}
 	}
 
+	FreeAccessStrategy(bas);
+
 	return stats;
 }
 
@@ -142,6 +142,11 @@ ivfflatvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
 {
 	Relation	rel = info->index;
 
+	if (info->analyze_only)
+		return stats;
+
+	/* stats is NULL if ambulkdelete not called */
+	/* OK to return NULL if index not changed */
 	if (stats == NULL)
 		return NULL;
 

src/vector.h

@@ -1,9 +1,7 @@
 #ifndef VECTOR_H
 #define VECTOR_H
 
-#include "postgres.h"
-
-#define VECTOR_MAX_DIM 1024
+#define VECTOR_MAX_DIM 16000
 
 #define VECTOR_SIZE(_dim)		(offsetof(Vector, x) + sizeof(float)*(_dim))
 #define DatumGetVector(x)		((Vector *) PG_DETOAST_DATUM(x))
@@ -18,24 +16,8 @@ typedef struct Vector
 	float		x[FLEXIBLE_ARRAY_MEMBER];
 }			Vector;
 
+Vector	   *InitVector(int dim);
 void		PrintVector(char *msg, Vector * vector);
 int			vector_cmp_internal(Vector * a, Vector * b);
 
-/*
- * Allocate and initialize a new vector
- */
-static inline Vector *
-InitVector(int dim)
-{
-	Vector	   *result;
-	int			size;
-
-	size = VECTOR_SIZE(dim);
-	result = (Vector *) palloc0(size);
-	SET_VARSIZE(result, size);
-	result->dim = dim;
-
-	return result;
-}
-
 #endif

test/expected/btree.out

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);

test/expected/cast.out

@@ -0,0 +1,59 @@
+SELECT ARRAY[1,2,3]::vector;
+  array  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT ARRAY[1.0,2.0,3.0]::vector;
+  array  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT ARRAY[1,2,3]::float4[]::vector;
+  array  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT ARRAY[1,2,3]::float8[]::vector;
+  array  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT ARRAY[1,2,3]::numeric[]::vector;
+  array  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT '{NULL}'::real[]::vector;
+ERROR:  array must not contain nulls
+SELECT '{NaN}'::real[]::vector;
+ERROR:  NaN not allowed in vector
+SELECT '{Infinity}'::real[]::vector;
+ERROR:  infinite value not allowed in vector
+SELECT '{-Infinity}'::real[]::vector;
+ERROR:  infinite value not allowed in vector
+SELECT '{}'::real[]::vector;
+ERROR:  vector must have at least 1 dimension
+SELECT '{{1}}'::real[]::vector;
+ERROR:  array must be 1-D
+SELECT '[1,2,3]'::vector::real[];
+ float4  
+---------
+ {1,2,3}
+(1 row)
+
+SELECT array_agg(n)::vector FROM generate_series(1, 16001) n;
+ERROR:  vector cannot have more than 16000 dimensions
+SELECT array_to_vector(array_agg(n), 16001, false) FROM generate_series(1, 16001) n;
+ERROR:  vector cannot have more than 16000 dimensions
+-- ensure no error
+SELECT ARRAY[1,2,3] = ARRAY[1,2,3];
+ ?column? 
+----------
+ t
+(1 row)
+

test/expected/copy.out

@@ -1,10 +1,8 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE TABLE t2 (val vector(3));
-\copy t TO '/tmp/data.bin' WITH (FORMAT binary)
-\copy t2 FROM '/tmp/data.bin' WITH (FORMAT binary)
+\copy t TO 'results/data.bin' WITH (FORMAT binary)
+\copy t2 FROM 'results/data.bin' WITH (FORMAT binary)
 SELECT * FROM t2 ORDER BY val;
    val   
 ---------

test/expected/functions.out

@@ -0,0 +1,210 @@
+SELECT '[1,2,3]'::vector + '[4,5,6]';
+ ?column? 
+----------
+ [5,7,9]
+(1 row)
+
+SELECT '[3e38]'::vector + '[3e38]';
+ERROR:  value out of range: overflow
+SELECT '[1,2,3]'::vector - '[4,5,6]';
+  ?column?  
+------------
+ [-3,-3,-3]
+(1 row)
+
+SELECT '[-3e38]'::vector - '[3e38]';
+ERROR:  value out of range: overflow
+SELECT '[1,2,3]'::vector * '[4,5,6]';
+ ?column?  
+-----------
+ [4,10,18]
+(1 row)
+
+SELECT '[1e37]'::vector * '[1e37]';
+ERROR:  value out of range: overflow
+SELECT '[1e-37]'::vector * '[1e-37]';
+ERROR:  value out of range: underflow
+SELECT vector_dims('[1,2,3]');
+ vector_dims 
+-------------
+           3
+(1 row)
+
+SELECT round(vector_norm('[1,1]')::numeric, 5);
+  round  
+---------
+ 1.41421
+(1 row)
+
+SELECT vector_norm('[3,4]');
+ vector_norm 
+-------------
+           5
+(1 row)
+
+SELECT vector_norm('[0,1]');
+ vector_norm 
+-------------
+           1
+(1 row)
+
+SELECT vector_norm('[3e37,4e37]')::real;
+ vector_norm 
+-------------
+       5e+37
+(1 row)
+
+SELECT l2_distance('[0,0]', '[3,4]');
+ l2_distance 
+-------------
+           5
+(1 row)
+
+SELECT l2_distance('[0,0]', '[0,1]');
+ l2_distance 
+-------------
+           1
+(1 row)
+
+SELECT l2_distance('[1,2]', '[3]');
+ERROR:  different vector dimensions 2 and 1
+SELECT l2_distance('[3e38]', '[-3e38]');
+ l2_distance 
+-------------
+    Infinity
+(1 row)
+
+SELECT inner_product('[1,2]', '[3,4]');
+ inner_product 
+---------------
+            11
+(1 row)
+
+SELECT inner_product('[1,2]', '[3]');
+ERROR:  different vector dimensions 2 and 1
+SELECT inner_product('[3e38]', '[3e38]');
+ inner_product 
+---------------
+      Infinity
+(1 row)
+
+SELECT cosine_distance('[1,2]', '[2,4]');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('[1,2]', '[0,0]');
+ cosine_distance 
+-----------------
+             NaN
+(1 row)
+
+SELECT cosine_distance('[1,1]', '[1,1]');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('[1,0]', '[0,2]');
+ cosine_distance 
+-----------------
+               1
+(1 row)
+
+SELECT cosine_distance('[1,1]', '[-1,-1]');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
+SELECT cosine_distance('[1,2]', '[3]');
+ERROR:  different vector dimensions 2 and 1
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
+SELECT cosine_distance('[3e38]', '[3e38]');
+ cosine_distance 
+-----------------
+             NaN
+(1 row)
+
+SELECT l1_distance('[0,0]', '[3,4]');
+ l1_distance 
+-------------
+           7
+(1 row)
+
+SELECT l1_distance('[0,0]', '[0,1]');
+ l1_distance 
+-------------
+           1
+(1 row)
+
+SELECT l1_distance('[1,2]', '[3]');
+ERROR:  different vector dimensions 2 and 1
+SELECT l1_distance('[3e38]', '[-3e38]');
+ l1_distance 
+-------------
+    Infinity
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
+    avg    
+-----------
+ [2,3.5,5]
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
+    avg    
+-----------
+ [2,3.5,5]
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
+ avg 
+-----
+ 
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
+ERROR:  expected 2 dimensions, not 1
+SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
+   avg   
+---------
+ [3e+38]
+(1 row)
+
+SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
+ERROR:  vector cannot have more than 16000 dimensions
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
+   sum    
+----------
+ [4,7,10]
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
+   sum    
+----------
+ [4,7,10]
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
+ sum 
+-----
+ 
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
+ERROR:  different vector dimensions 2 and 1
+SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
+ERROR:  value out of range: overflow

test/expected/hnsw_cosine.out

@@ -0,0 +1,26 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_cosine_ops);
+INSERT INTO t (val) VALUES ('[1,2,4]');
+SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+   val   
+---------
+ [1,1,1]
+ [1,2,3]
+ [1,2,4]
+(3 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+ count 
+-------
+     3
+(1 row)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+ count 
+-------
+     3
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_ip.out

@@ -0,0 +1,21 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_ip_ops);
+INSERT INTO t (val) VALUES ('[1,2,4]');
+SELECT * FROM t ORDER BY val <#> '[3,3,3]';
+   val   
+---------
+ [1,2,4]
+ [1,2,3]
+ [1,1,1]
+ [0,0,0]
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_l2.out

@@ -1,9 +1,7 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
    val   
@@ -15,6 +13,22 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
 (4 rows)
 
 SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+   val   
+---------
+ [0,0,0]
+ [1,1,1]
+ [1,2,3]
+ [1,2,4]
+(4 rows)
+
+SELECT COUNT(*) FROM t;
+ count 
+-------
+     5
+(1 row)
+
+TRUNCATE t;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
  val 
 -----
 (0 rows)

test/expected/hnsw_options.out

@@ -0,0 +1,26 @@
+CREATE TABLE t (val vector(3));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
+ERROR:  value 1 out of bounds for option "m"
+DETAIL:  Valid values are between "2" and "100".
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
+ERROR:  value 101 out of bounds for option "m"
+DETAIL:  Valid values are between "2" and "100".
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
+ERROR:  value 3 out of bounds for option "ef_construction"
+DETAIL:  Valid values are between "4" and "1000".
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
+ERROR:  value 1001 out of bounds for option "ef_construction"
+DETAIL:  Valid values are between "4" and "1000".
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
+ERROR:  ef_construction must be greater than or equal to 2 * m
+SHOW hnsw.ef_search;
+ hnsw.ef_search 
+----------------
+ 40
+(1 row)
+
+SET hnsw.ef_search = 0;
+ERROR:  0 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
+SET hnsw.ef_search = 1001;
+ERROR:  1001 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
+DROP TABLE t;

test/expected/hnsw_unlogged.out

@@ -0,0 +1,13 @@
+SET enable_seqscan = off;
+CREATE UNLOGGED TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,1,1]
+ [0,0,0]
+(3 rows)
+
+DROP TABLE t;

test/expected/input.out

@@ -0,0 +1,129 @@
+SELECT '[1,2,3]'::vector;
+ vector  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT '[-1,-2,-3]'::vector;
+   vector   
+------------
+ [-1,-2,-3]
+(1 row)
+
+SELECT '[1.,2.,3.]'::vector;
+ vector  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT ' [ 1,  2 ,    3  ] '::vector;
+ vector  
+---------
+ [1,2,3]
+(1 row)
+
+SELECT '[1.23456]'::vector;
+  vector   
+-----------
+ [1.23456]
+(1 row)
+
+SELECT '[hello,1]'::vector;
+ERROR:  invalid input syntax for type vector: "[hello,1]"
+LINE 1: SELECT '[hello,1]'::vector;
+               ^
+SELECT '[NaN,1]'::vector;
+ERROR:  NaN not allowed in vector
+LINE 1: SELECT '[NaN,1]'::vector;
+               ^
+SELECT '[Infinity,1]'::vector;
+ERROR:  infinite value not allowed in vector
+LINE 1: SELECT '[Infinity,1]'::vector;
+               ^
+SELECT '[-Infinity,1]'::vector;
+ERROR:  infinite value not allowed in vector
+LINE 1: SELECT '[-Infinity,1]'::vector;
+               ^
+SELECT '[1.5e38,-1.5e38]'::vector;
+       vector       
+--------------------
+ [1.5e+38,-1.5e+38]
+(1 row)
+
+SELECT '[1.5e+38,-1.5e+38]'::vector;
+       vector       
+--------------------
+ [1.5e+38,-1.5e+38]
+(1 row)
+
+SELECT '[1.5e-38,-1.5e-38]'::vector;
+       vector       
+--------------------
+ [1.5e-38,-1.5e-38]
+(1 row)
+
+SELECT '[4e38,1]'::vector;
+ERROR:  infinite value not allowed in vector
+LINE 1: SELECT '[4e38,1]'::vector;
+               ^
+SELECT '[1,2,3'::vector;
+ERROR:  malformed vector literal: "[1,2,3"
+LINE 1: SELECT '[1,2,3'::vector;
+               ^
+DETAIL:  Unexpected end of input.
+SELECT '[1,2,3]9'::vector;
+ERROR:  malformed vector literal: "[1,2,3]9"
+LINE 1: SELECT '[1,2,3]9'::vector;
+               ^
+DETAIL:  Junk after closing right brace.
+SELECT '1,2,3'::vector;
+ERROR:  malformed vector literal: "1,2,3"
+LINE 1: SELECT '1,2,3'::vector;
+               ^
+DETAIL:  Vector contents must start with "[".
+SELECT ''::vector;
+ERROR:  malformed vector literal: ""
+LINE 1: SELECT ''::vector;
+               ^
+DETAIL:  Vector contents must start with "[".
+SELECT '['::vector;
+ERROR:  malformed vector literal: "["
+LINE 1: SELECT '['::vector;
+               ^
+DETAIL:  Unexpected end of input.
+SELECT '[,'::vector;
+ERROR:  malformed vector literal: "[,"
+LINE 1: SELECT '[,'::vector;
+               ^
+DETAIL:  Unexpected end of input.
+SELECT '[]'::vector;
+ERROR:  vector must have at least 1 dimension
+LINE 1: SELECT '[]'::vector;
+               ^
+SELECT '[1,]'::vector;
+ERROR:  invalid input syntax for type vector: "[1,]"
+LINE 1: SELECT '[1,]'::vector;
+               ^
+SELECT '[1a]'::vector;
+ERROR:  invalid input syntax for type vector: "[1a]"
+LINE 1: SELECT '[1a]'::vector;
+               ^
+SELECT '[1,,3]'::vector;
+ERROR:  malformed vector literal: "[1,,3]"
+LINE 1: SELECT '[1,,3]'::vector;
+               ^
+SELECT '[1, ,3]'::vector;
+ERROR:  invalid input syntax for type vector: "[1, ,3]"
+LINE 1: SELECT '[1, ,3]'::vector;
+               ^
+SELECT '[1,2,3]'::vector(2);
+ERROR:  expected 2 dimensions, not 3
+SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::vector[]);
+ unnest  
+---------
+ [1,2,3]
+ [4,5,6]
+(2 rows)
+
+SELECT '{"[1,2,3]"}'::vector(2)[];
+ERROR:  expected 2 dimensions, not 3

test/expected/ivfflat_cosine.out

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
@@ -13,9 +11,16 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,2,4]
 (3 rows)
 
-SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector);
- val 
------
-(0 rows)
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+ count 
+-------
+     3
+(1 row)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+ count 
+-------
+     3
+(1 row)
 
 DROP TABLE t;

test/expected/ivfflat_ip.out

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
@@ -14,9 +12,10 @@ SELECT * FROM t ORDER BY val <#> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector);
- val 
------
-(0 rows)
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+ count 
+-------
+     4
+(1 row)
 
 DROP TABLE t;

test/expected/ivfflat_l2.out

@@ -0,0 +1,39 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
+INSERT INTO t (val) VALUES ('[1,2,4]');
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,2,4]
+ [1,1,1]
+ [0,0,0]
+(4 rows)
+
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+   val   
+---------
+ [0,0,0]
+ [1,1,1]
+ [1,2,3]
+ [1,2,4]
+(4 rows)
+
+SELECT COUNT(*) FROM t;
+ count 
+-------
+     5
+(1 row)
+
+TRUNCATE t;
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+ val 
+-----
+(0 rows)
+
+DROP TABLE t;

test/expected/ivfflat_options.out

@@ -0,0 +1,14 @@
+CREATE TABLE t (val vector(3));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
+ERROR:  value 0 out of bounds for option "lists"
+DETAIL:  Valid values are between "1" and "32768".
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
+ERROR:  value 32769 out of bounds for option "lists"
+DETAIL:  Valid values are between "1" and "32768".
+SHOW ivfflat.probes;
+ ivfflat.probes 
+----------------
+ 1
+(1 row)
+
+DROP TABLE t;

test/expected/ivfflat_unlogged.out

@@ -1,9 +1,7 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 CREATE UNLOGGED TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
    val   
 ---------

test/perl/PostgresNode.pm

@@ -0,0 +1,8 @@
+use PostgreSQL::Test::Cluster;
+
+sub get_new_node
+{
+	return PostgreSQL::Test::Cluster->new(@_);
+}
+
+1;

test/perl/TestLib.pm

@@ -0,0 +1,3 @@
+use PostgreSQL::Test::Utils;
+
+1;

test/sql/btree.sql

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 
 CREATE TABLE t (val vector(3));

test/sql/cast.sql

@@ -0,0 +1,17 @@
+SELECT ARRAY[1,2,3]::vector;
+SELECT ARRAY[1.0,2.0,3.0]::vector;
+SELECT ARRAY[1,2,3]::float4[]::vector;
+SELECT ARRAY[1,2,3]::float8[]::vector;
+SELECT ARRAY[1,2,3]::numeric[]::vector;
+SELECT '{NULL}'::real[]::vector;
+SELECT '{NaN}'::real[]::vector;
+SELECT '{Infinity}'::real[]::vector;
+SELECT '{-Infinity}'::real[]::vector;
+SELECT '{}'::real[]::vector;
+SELECT '{{1}}'::real[]::vector;
+SELECT '[1,2,3]'::vector::real[];
+SELECT array_agg(n)::vector FROM generate_series(1, 16001) n;
+SELECT array_to_vector(array_agg(n), 16001, false) FROM generate_series(1, 16001) n;
+
+-- ensure no error
+SELECT ARRAY[1,2,3] = ARRAY[1,2,3];

test/sql/copy.sql

@@ -1,13 +1,10 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 
 CREATE TABLE t2 (val vector(3));
 
-\copy t TO '/tmp/data.bin' WITH (FORMAT binary)
-\copy t2 FROM '/tmp/data.bin' WITH (FORMAT binary)
+\copy t TO 'results/data.bin' WITH (FORMAT binary)
+\copy t2 FROM 'results/data.bin' WITH (FORMAT binary)
 
 SELECT * FROM t2 ORDER BY val;
 

test/sql/functions.sql

@@ -0,0 +1,51 @@
+SELECT '[1,2,3]'::vector + '[4,5,6]';
+SELECT '[3e38]'::vector + '[3e38]';
+SELECT '[1,2,3]'::vector - '[4,5,6]';
+SELECT '[-3e38]'::vector - '[3e38]';
+SELECT '[1,2,3]'::vector * '[4,5,6]';
+SELECT '[1e37]'::vector * '[1e37]';
+SELECT '[1e-37]'::vector * '[1e-37]';
+
+SELECT vector_dims('[1,2,3]');
+
+SELECT round(vector_norm('[1,1]')::numeric, 5);
+SELECT vector_norm('[3,4]');
+SELECT vector_norm('[0,1]');
+SELECT vector_norm('[3e37,4e37]')::real;
+
+SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]', '[0,1]');
+SELECT l2_distance('[1,2]', '[3]');
+SELECT l2_distance('[3e38]', '[-3e38]');
+
+SELECT inner_product('[1,2]', '[3,4]');
+SELECT inner_product('[1,2]', '[3]');
+SELECT inner_product('[3e38]', '[3e38]');
+
+SELECT cosine_distance('[1,2]', '[2,4]');
+SELECT cosine_distance('[1,2]', '[0,0]');
+SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,0]', '[0,2]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
+SELECT cosine_distance('[1,2]', '[3]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
+SELECT cosine_distance('[3e38]', '[3e38]');
+
+SELECT l1_distance('[0,0]', '[3,4]');
+SELECT l1_distance('[0,0]', '[0,1]');
+SELECT l1_distance('[1,2]', '[3]');
+SELECT l1_distance('[3e38]', '[-3e38]');
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
+SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
+SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
+SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
+SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
+SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
+SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
+SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;

test/sql/hnsw_cosine.sql

@@ -0,0 +1,13 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_cosine_ops);
+
+INSERT INTO t (val) VALUES ('[1,2,4]');
+
+SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+
+DROP TABLE t;

test/sql/hnsw_ip.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_ip_ops);
+
+INSERT INTO t (val) VALUES ('[1,2,4]');
+
+SELECT * FROM t ORDER BY val <#> '[3,3,3]';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+
+DROP TABLE t;

test/sql/hnsw_l2.sql

@@ -1,14 +1,16 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 
 CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
 SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT COUNT(*) FROM t;
+
+TRUNCATE t;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
 
 DROP TABLE t;

test/sql/hnsw_options.sql

@@ -0,0 +1,13 @@
+CREATE TABLE t (val vector(3));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
+
+SHOW hnsw.ef_search;
+
+SET hnsw.ef_search = 0;
+SET hnsw.ef_search = 1001;
+
+DROP TABLE t;

test/sql/hnsw_unlogged.sql

@@ -0,0 +1,9 @@
+SET enable_seqscan = off;
+
+CREATE UNLOGGED TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+DROP TABLE t;

test/sql/input.sql

@@ -0,0 +1,28 @@
+SELECT '[1,2,3]'::vector;
+SELECT '[-1,-2,-3]'::vector;
+SELECT '[1.,2.,3.]'::vector;
+SELECT ' [ 1,  2 ,    3  ] '::vector;
+SELECT '[1.23456]'::vector;
+SELECT '[hello,1]'::vector;
+SELECT '[NaN,1]'::vector;
+SELECT '[Infinity,1]'::vector;
+SELECT '[-Infinity,1]'::vector;
+SELECT '[1.5e38,-1.5e38]'::vector;
+SELECT '[1.5e+38,-1.5e+38]'::vector;
+SELECT '[1.5e-38,-1.5e-38]'::vector;
+SELECT '[4e38,1]'::vector;
+SELECT '[1,2,3'::vector;
+SELECT '[1,2,3]9'::vector;
+SELECT '1,2,3'::vector;
+SELECT ''::vector;
+SELECT '['::vector;
+SELECT '[,'::vector;
+SELECT '[]'::vector;
+SELECT '[1,]'::vector;
+SELECT '[1a]'::vector;
+SELECT '[1,,3]'::vector;
+SELECT '[1, ,3]'::vector;
+SELECT '[1,2,3]'::vector(2);
+
+SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::vector[]);
+SELECT '{"[1,2,3]"}'::vector(2)[];

test/sql/ivfflat_cosine.sql

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 
 CREATE TABLE t (val vector(3));
@@ -9,6 +7,7 @@ CREATE INDEX ON t USING ivfflat (val vector_cosine_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
 SELECT * FROM t ORDER BY val <=> '[3,3,3]';
-SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector);
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
 
 DROP TABLE t;

test/sql/ivfflat_ip.sql

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 
 CREATE TABLE t (val vector(3));
@@ -9,6 +7,6 @@ CREATE INDEX ON t USING ivfflat (val vector_ip_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
 SELECT * FROM t ORDER BY val <#> '[3,3,3]';
-SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector);
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
 
 DROP TABLE t;

test/sql/ivfflat_l2.sql

@@ -0,0 +1,16 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
+
+INSERT INTO t (val) VALUES ('[1,2,4]');
+
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT COUNT(*) FROM t;
+
+TRUNCATE t;
+SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+
+DROP TABLE t;

test/sql/ivfflat_options.sql

@@ -0,0 +1,7 @@
+CREATE TABLE t (val vector(3));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
+
+SHOW ivfflat.probes;
+
+DROP TABLE t;

test/sql/ivfflat_unlogged.sql

@@ -1,10 +1,8 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SET enable_seqscan = off;
 
 CREATE UNLOGGED TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
 
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
 

test/t/001_ivfflat_wal.pl

@@ -5,7 +5,9 @@ use strict;
 use warnings;
 use PostgresNode;
 use TestLib;
-use Test::More tests => 31;
+use Test::More;
+
+my $dim = 32;
 
 my $node_primary;
 my $node_replica;
@@ -17,26 +19,21 @@ sub test_index_replay
 
 	# Wait for replica to catch up
 	my $applname = $node_replica->name;
-
-	my $caughtup_query;
-	my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
-	if ($server_version_num >= 100000) {
-		$caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
-	} else {
-		# TODO figure out why replay location doesn't work
-		$caughtup_query = "SELECT pg_current_xlog_location() <= write_location FROM pg_stat_replication WHERE application_name = '$applname';";
-	}
-
+	my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
 	$node_primary->poll_query_until('postgres', $caughtup_query)
 	  or die "Timed out while waiting for replica 1 to catch up";
 
-	my $r1 = rand();
-	my $r2 = rand();
-	my $r3 = rand();
+	my @r = ();
+	for (1 .. $dim)
+	{
+		push(@r, rand());
+	}
+	my $sql = join(",", @r);
 
-	my $queries = qq(SET enable_seqscan=off;
-SELECT * FROM tst ORDER BY v <-> '[$r1,$r2,$r3]' LIMIT 10;
-);
+	my $queries = qq(
+		SET enable_seqscan = off;
+		SELECT * FROM tst ORDER BY v <-> '[$sql]' LIMIT 10;
+	);
 
 	# Run test queries and compare their result
 	my $primary_result = $node_primary->safe_psql("postgres", $queries);
@@ -46,9 +43,23 @@ SELECT * FROM tst ORDER BY v <-> '[$r1,$r2,$r3]' LIMIT 10;
 	return;
 }
 
+# Use ARRAY[random(), random(), random(), ...] over
+# SELECT array_agg(random()) FROM generate_series(1, $dim)
+# to generate different values for each row
+my $array_sql = join(",", ('random()') x $dim);
+
 # Initialize primary node
 $node_primary = get_new_node('primary');
 $node_primary->init(allows_streaming => 1);
+if ($dim > 32)
+{
+	# TODO use wal_keep_segments for Postgres < 13
+	$node_primary->append_conf('postgresql.conf', qq(wal_keep_size = 1GB));
+}
+if ($dim > 1500)
+{
+	$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
+}
 $node_primary->start;
 my $backup_name = 'my_backup';
 
@@ -57,18 +68,16 @@ $node_primary->backup($backup_name);
 
 # Create streaming replica linking to primary
 $node_replica = get_new_node('replica');
-$node_replica->init_from_backup($node_primary, $backup_name,
-	has_streaming => 1);
+$node_replica->init_from_backup($node_primary, $backup_name, has_streaming => 1);
 $node_replica->start;
 
 # Create ivfflat index on primary
 $node_primary->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node_primary->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
+$node_primary->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node_primary->safe_psql("postgres",
-	"INSERT INTO tst SELECT i%10, ARRAY[random(), random(), random()] FROM generate_series(1,100000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
-$node_primary->safe_psql("postgres",
-	"CREATE INDEX ON tst USING ivfflat (v);");
+$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
 
 # Test that queries give same result
 test_index_replay('initial');
@@ -82,7 +91,9 @@ for my $i (1 .. 10)
 	test_index_replay("vacuum $i");
 	my ($start, $end) = (100001 + ($i - 1) * 10000, 100000 + $i * 10000);
 	$node_primary->safe_psql("postgres",
-		"INSERT INTO tst SELECT i%10, ARRAY[random(), random(), random()] FROM generate_series($start,$end) i;"
+		"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series($start, $end) i;"
 	);
 	test_index_replay("insert $i");
 }
+
+done_testing();

test/t/002_ivfflat_vacuum.pl

@@ -0,0 +1,44 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $dim = 3;
+
+my @r = ();
+for (1 .. $dim)
+{
+	my $v = int(rand(1000)) + 1;
+	push(@r, "i % $v");
+}
+my $array_sql = join(", ", @r);
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+
+# Get size
+my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
+
+# Delete all, vacuum, and insert same data
+$node->safe_psql("postgres", "DELETE FROM tst;");
+$node->safe_psql("postgres", "VACUUM tst;");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
+);
+
+# Check size
+my $new_size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
+is($size, $new_size, "size does not change");
+
+done_testing();

test/t/003_ivfflat_build_recall.pl

@@ -0,0 +1,128 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+
+sub test_recall
+{
+	my ($probes, $min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SET ivfflat.probes = $probes;
+		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Scan using idx on tst/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SET ivfflat.probes = $probes;
+			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
+		));
+		my @actual_ids = split("\n", $actual);
+		my %actual_set = map { $_ => 1 } @actual_ids;
+
+		my @expected_ids = split("\n", $expected[$i]);
+
+		foreach (@expected_ids)
+		{
+			if (exists($actual_set{$_}))
+			{
+				$correct++;
+			}
+			$total++;
+		}
+	}
+
+	cmp_ok($correct / $total, ">=", $min, $operator);
+}
+
+# Initialize node
+$node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "[$r1,$r2,$r3]");
+}
+
+# Check each index type
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	# Get exact results
+	@expected = ();
+	foreach (@queries)
+	{
+		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
+		push(@expected, $res);
+	}
+
+	# Build index serially
+	$node->safe_psql("postgres", qq(
+		SET max_parallel_maintenance_workers = 0;
+		CREATE INDEX idx ON tst USING ivfflat (v $opclass);
+	));
+
+	# Test approximate results
+	if ($operator ne "<#>")
+	{
+		# TODO Fix test (uniform random vectors all have similar inner product)
+		test_recall(1, 0.71, $operator);
+		test_recall(10, 0.95, $operator);
+	}
+	# Account for equal distances
+	test_recall(100, 0.9925, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+
+	# Build index in parallel
+	my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+		SET client_min_messages = DEBUG;
+		SET min_parallel_table_scan_size = 1;
+		CREATE INDEX idx ON tst USING ivfflat (v $opclass);
+	));
+	is($ret, 0, $stderr);
+	like($stderr, qr/using \d+ parallel workers/);
+
+	# Test approximate results
+	if ($operator ne "<#>")
+	{
+		# TODO Fix test (uniform random vectors all have similar inner product)
+		test_recall(1, 0.71, $operator);
+		test_recall(10, 0.95, $operator);
+	}
+	# Account for equal distances
+	test_recall(100, 0.9925, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();

test/t/004_ivfflat_centers.pl

@@ -0,0 +1,38 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, '[1,2,3]' FROM generate_series(1, 10) i;"
+);
+
+sub test_centers
+{
+	my ($lists, $min) = @_;
+
+	my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops) WITH (lists = $lists);");
+	is($ret, 0, $stderr);
+}
+
+# Test no error for duplicate centers
+test_centers(5);
+test_centers(10);
+
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, '[4,5,6]' FROM generate_series(1, 10) i;"
+);
+
+# Test no error for duplicate centers
+test_centers(10);
+
+done_testing();

test/t/005_ivfflat_query_recall.pl

@@ -0,0 +1,44 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4 primary key, v vector(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
+);
+
+# Check each index type
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v $opclass);");
+
+	# Test 100% recall
+	for (1 .. 20)
+	{
+		my $id = int(rand() * 100000);
+		my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $id;");
+		my $res = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SELECT v FROM tst ORDER BY v <-> '$query' LIMIT 1;
+		));
+		is($res, $query);
+	}
+}
+
+done_testing();

test/t/006_ivfflat_lists.pl

@@ -0,0 +1,33 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (v vector(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
+);
+
+$node->safe_psql("postgres", "CREATE INDEX lists50 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 50);");
+$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 100);");
+
+# Test prefers more lists
+my $res = $node->safe_psql("postgres", "EXPLAIN SELECT v FROM tst ORDER BY v <-> '[0.5,0.5,0.5]' LIMIT 10;");
+like($res, qr/lists100/);
+unlike($res, qr/lists50/);
+
+# Test errors with too much memory
+my ($ret, $stdout, $stderr) = $node->psql("postgres",
+	"CREATE INDEX lists10000 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10000);"
+);
+like($stderr, qr/memory required is/);
+
+done_testing();

test/t/007_ivfflat_inserts.pl

@@ -0,0 +1,57 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $dim = 768;
+
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+
+$node->pgbench(
+	"--no-vacuum --client=5 --transactions=100",
+	0,
+	[qr{actually processed}],
+	[qr{^$}],
+	"concurrent INSERTs",
+	{
+		"007_ivfflat_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
+	}
+);
+
+sub idx_scan
+{
+	# Stats do not update instantaneously
+	# https://www.postgresql.org/docs/current/monitoring-stats.html#MONITORING-STATS-VIEWS
+	sleep(1);
+	$node->safe_psql("postgres", "SELECT idx_scan FROM pg_stat_user_indexes WHERE indexrelid = 'tst_v_idx'::regclass;");
+}
+
+my $expected = 10000 + 5 * 100 * 10;
+
+my $count = $node->safe_psql("postgres", "SELECT COUNT(*) FROM tst;");
+is($count, $expected);
+is(idx_scan(), 0);
+
+$count = $node->safe_psql("postgres", qq(
+	SET enable_seqscan = off;
+	SET ivfflat.probes = 100;
+	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
+));
+is($count, $expected);
+is(idx_scan(), 1);
+
+done_testing();

test/t/008_aggregates.pl

@@ -0,0 +1,49 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (r1 real, r2 real, r3 real, v vector(3));");
+$node->safe_psql("postgres", qq(
+	INSERT INTO tst SELECT r1, r2, r3, ARRAY[r1, r2, r3] FROM (
+		SELECT random() + 1.01 AS r1, random() + 2.01 AS r2, random() + 3.01 AS r3 FROM generate_series(1, 1000000) t
+	) i;
+));
+
+sub test_aggregate
+{
+	my ($agg) = @_;
+
+	# Test value
+	my $res = $node->safe_psql("postgres", "SELECT $agg(v) FROM tst;");
+	like($res, qr/\[1\.5/);
+	like($res, qr/,2\.5/);
+	like($res, qr/,3\.5/);
+
+	# Test matches real for avg
+	# Cannot test sum since sum(real) varies between calls
+	if ($agg eq 'avg')
+	{
+		my $r1 = $node->safe_psql("postgres", "SELECT $agg(r1)::float4 FROM tst;");
+		my $r2 = $node->safe_psql("postgres", "SELECT $agg(r2)::float4 FROM tst;");
+		my $r3 = $node->safe_psql("postgres", "SELECT $agg(r3)::float4 FROM tst;");
+		is($res, "[$r1,$r2,$r3]");
+	}
+
+	# Test explain
+	my $explain = $node->safe_psql("postgres", "EXPLAIN SELECT $agg(v) FROM tst;");
+	like($explain, qr/Partial Aggregate/);
+}
+
+test_aggregate('avg');
+test_aggregate('sum');
+
+done_testing();

test/t/009_storage.pl

@@ -0,0 +1,34 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $dim = 1024;
+
+# Initialize node
+my $node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (v1 vector(1024), v2 vector(1024), v3 vector(1024));");
+
+# Test insert succeeds
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT array_agg(n), array_agg(n), array_agg(n) FROM generate_series(1, $dim) n"
+);
+
+# Change storage to PLAIN
+$node->safe_psql("postgres", "ALTER TABLE tst ALTER COLUMN v1 SET STORAGE PLAIN");
+$node->safe_psql("postgres", "ALTER TABLE tst ALTER COLUMN v2 SET STORAGE PLAIN");
+$node->safe_psql("postgres", "ALTER TABLE tst ALTER COLUMN v3 SET STORAGE PLAIN");
+
+# Test insert fails
+my ($ret, $stdout, $stderr) = $node->psql("postgres",
+	"INSERT INTO tst SELECT array_agg(n), array_agg(n), array_agg(n) FROM generate_series(1, $dim) n"
+);
+like($stderr, qr/row is too big/);
+
+done_testing();