Added query-aware dynamic pruning [skip ci]

.editorconfig

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

.github/workflows/build.yml

@@ -1,39 +1,72 @@
 name: build
 on: [push, pull_request]
 jobs:
-  build:
+  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: [15, 14, 13, 12, 11, 10]
         include:
-        - os: macos-latest
-          postgres: 14
+          - 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@v3
-    - uses: ankane/setup-postgres@v1
-      with:
-        postgres-version: ${{ matrix.postgres }}
-        dev-files: true
-    - if: ${{ startsWith(matrix.os, 'ubuntu') }}
-      run: sudo apt-get update && sudo apt-get install 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_14_5.tar.gz
-        tar xf REL_14_5.tar.gz
-        make prove_installcheck PROVE=prove PROVE_FLAGS="-I ./postgres-REL_14_5/src/test/perl" PERL5LIB="/Users/runner/perl5/lib/perl5"
+      - uses: actions/checkout@v3
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: ${{ matrix.postgres }}
+          dev-files: true
+      - run: make
+      - 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
+      - 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"
+  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

.gitignore

@@ -1,4 +1,5 @@
 /dist/
+/log/
 /results/
 /tmp_check/
 /sql/vector--?.?.?.sql
@@ -6,3 +7,8 @@ regression.*
 *.o
 *.so
 *.bc
+*.dll
+*.dylib
+*.obj
+*.lib
+*.exp

CHANGELOG.md

@@ -1,3 +1,26 @@
+## 0.4.2 (unreleased)
+
+- Added notice when index created with little data
+- 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

Dockerfile

@@ -1,9 +1,11 @@
-FROM postgres:15
+ARG PG_MAJOR=15
+FROM postgres:$PG_MAJOR
+ARG PG_MAJOR
 
 COPY . /tmp/pgvector
 
 RUN apt-get update && \
-		apt-get install -y --no-install-recommends build-essential postgresql-server-dev-15 && \
+		apt-get install -y --no-install-recommends build-essential postgresql-server-dev-$PG_MAJOR && \
 		cd /tmp/pgvector && \
 		make clean && \
 		make OPTFLAGS="" && \
@@ -11,6 +13,6 @@ RUN apt-get update && \
 		mkdir /usr/share/doc/pgvector && \
 		cp LICENSE README.md /usr/share/doc/pgvector && \
 		rm -r /tmp/pgvector && \
-		apt-get remove -y build-essential postgresql-server-dev-15 && \
+		apt-get remove -y build-essential postgresql-server-dev-$PG_MAJOR && \
 		apt-get autoremove -y && \
 		rm -rf /var/lib/apt/lists/*

LICENSE

@@ -1,4 +1,4 @@
-Portions Copyright (c) 1996-2021, PostgreSQL Global Development Group
+Portions Copyright (c) 1996-2022, PostgreSQL Global Development Group
 
 Portions Copyright (c) 1994, The Regents of the University of California
 

META.json

@@ -2,7 +2,7 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.3.2",
+	"version": "0.4.1",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -12,7 +12,7 @@
 	"prereqs": {
 		"runtime": {
 			"requires": {
-				"PostgreSQL": "10.0.0"
+				"PostgreSQL": "11.0.0"
 			}
 		}
 	},
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.3.2",
+			"version": "0.4.1",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.3.2
+EXTVERSION = 0.4.1
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
@@ -7,13 +7,14 @@ OBJS = src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
-REGRESS_OPTS = --inputdir=test
+REGRESS_OPTS = --inputdir=test --load-extension=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
@@ -40,6 +41,11 @@ 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
 
@@ -56,4 +62,4 @@ dist:
 .PHONY: docker
 
 docker:
-	docker build --pull --no-cache -t ankane/pgvector:latest .
+	docker build --pull --no-cache --platform linux/amd64 -t ankane/pgvector:latest .

Makefile.win

@@ -0,0 +1,70 @@
+EXTENSION = vector
+EXTVERSION = 0.4.1
+
+OBJS = src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
+
+REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
+REGRESS_OPTS = --inputdir=test --load-extension=vector
+
+# 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 "$(SHAREDIR)\extension"
+
+installcheck:
+	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
+
+uninstall:
+	del /f "$(PKGLIBDIR)\$(SHLIB)"
+	del /f "$(SHAREDIR)\extension\$(EXTENSION).control"
+	del /f "$(SHAREDIR)\extension\vector--*.sql"
+
+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

@@ -2,22 +2,21 @@
 
 Open-source vector similarity search for Postgres
 
-```sql
-CREATE TABLE items (embedding vector(3));
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops);
-SELECT * FROM items ORDER BY embedding <-> '[1,2,3]' LIMIT 5;
-```
+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/pgvector/pgvector/workflows/build/badge.svg?branch=master)](https://github.com/pgvector/pgvector/actions)
 
 ## Installation
 
-Compile and install the extension (supports Postgres 10+)
+Compile and install the extension (supports Postgres 11+)
 
 ```sh
-git clone --branch v0.3.2 https://github.com/pgvector/pgvector.git
+cd /tmp
+git clone --branch v0.4.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -29,81 +28,178 @@ Then load it in databases where you want to use it
 CREATE EXTENSION vector;
 ```
 
-You can also install it with [Docker](#docker), [Homebrew](#homebrew), or [PGXN](#pgxn)
+See the [installation notes](#installation-notes) if you run into issues
+
+You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [Yum](#yum), or [conda-forge](#conda-forge)
 
 ## Getting Started
 
 Create a vector column with 3 dimensions
 
 ```sql
-CREATE TABLE items (embedding vector(3));
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
 ```
 
-Insert values
+Insert vectors
 
 ```sql
-INSERT INTO items VALUES ('[1,2,3]'), ('[4,5,6]');
+INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
 ```
 
-Get the nearest neighbor by L2 distance
+Get the nearest neighbors by L2 distance
 
 ```sql
-SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 1;
+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)
+
+```sql
+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 performance. Unlike typical indexes, you will see different results for queries after adding an approximate index.
+
+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 `lists / 10` for up to 1M rows and `sqrt(lists)` for over 1M rows
+
+Add an index for each distance function you want to use.
 
 L2 distance
 
-```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops);
-```
-
-Inner product
-
-```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_ip_ops);
-```
-
-Cosine distance
-
-```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_cosine_ops);
-```
-
-Indexes should be created after the table has some 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 items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100);
 ```
 
-A [good place to start](https://github.com/facebookresearch/faiss/issues/112) is `4 * sqrt(rows)`
+Inner product
+
+```sql
+CREATE INDEX ON items USING ivfflat (embedding vector_ip_ops) WITH (lists = 100);
+```
+
+Cosine distance
+
+```sql
+CREATE INDEX ON items USING ivfflat (embedding vector_cosine_ops) 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;
 ```
@@ -119,10 +215,9 @@ SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
 The phases are:
 
 1. `initializing`
-2. `sampling table`
-3. `performing k-means`
-4. `sorting tuples`
-5. `loading tuples`
+2. `performing k-means`
+3. `sorting tuples`
+4. `loading tuples`
 
 Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
 
@@ -137,7 +232,7 @@ SELECT * FROM items WHERE category_id = 123 ORDER BY embedding <-> '[3,1,2]' LIM
 can be indexed with:
 
 ```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WHERE (category_id = 123);
+CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100) WHERE (category_id = 123);
 ```
 
 To index many different values of `category_id`, consider [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html) on `category_id`.
@@ -148,23 +243,84 @@ CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(cate
 
 ## 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.
+
+```sql
+SELECT * FROM items ORDER BY embedding <#> '[3,1,2]' LIMIT 5;
+```
+
+### Approximate Search
+
 To speed up queries with an 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)
+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?
+
+Two things you can try are:
+
+1. use dimensionality reduction
+2. compile Postgres with a larger block size (`./configure --with-blocksize=32`) and edit the limit in `src/ivfflat.h`
+
+#### Why am I seeing less results after adding an index?
+
+The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
+
 ## 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
 
@@ -180,42 +336,53 @@ Operator | Description
 
 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
+cosine_distance(vector, vector) → double precision | cosine distance
+inner_product(vector, vector) → double precision | inner product
+l2_distance(vector, vector) → double precision | Euclidean distance
+vector_dims(vector) → integer | number of dimensions
+vector_norm(vector) → double precision | Euclidean norm
 
-## Libraries
+### Aggregate Functions
 
-Libraries that use pgvector:
+Function | Description
+--- | ---
+avg(vector) → vector | arithmetic mean
 
-- [pgvector-python](https://github.com/pgvector/pgvector-python) (Python)
-- [Neighbor](https://github.com/ankane/neighbor) (Ruby)
-- [pgvector-ruby](https://github.com/pgvector/pgvector-ruby) (Ruby)
-- [pgvector-node](https://github.com/pgvector/pgvector-node) (Node.js)
-- [pgvector-go](https://github.com/pgvector/pgvector-go) (Go)
-- [pgvector-php](https://github.com/pgvector/pgvector-php) (PHP)
-- [pgvector-rust](https://github.com/pgvector/pgvector-rust) (Rust)
-- [pgvector-cpp](https://github.com/pgvector/pgvector-cpp) (C++)
-- [pgvector-elixir](https://github.com/pgvector/pgvector-elixir) (Elixir)
+## Installation Notes
 
-## Frequently Asked Questions
+### Postgres Location
 
-#### How many vectors can be stored in a single table?
+If your machine has multiple Postgres installations, specify the path to [pg_config](https://www.postgresql.org/docs/current/app-pgconfig.html) with:
 
-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.
+```sh
+export PG_CONFIG=/Applications/Postgres.app/Contents/Versions/latest/bin/pg_config
+```
 
-#### Is replication supported?
+Then re-run the installation instructions (run `make clean` before `make` if needed)
 
-Yes, pgvector uses the write-ahead log (WAL), which allows for replication and point-in-time recovery.
+### Missing Header
 
-#### What if my data has more than 1024 dimensions?
+If compilation fails with `fatal error: postgres.h: No such file or directory`, make sure Postgres development files are installed on the server.
 
-Two things you can try are:
+For Ubuntu and Debian, use:
 
-1. use dimensionality reduction
-2. compile Postgres with a larger block size (`./configure --with-blocksize=32`) and edit the limit in `src/vector.h`
+```sh
+sudo apt-get install postgresql-server-dev-15
+```
+
+Note: Replace `15` with your Postgres server version
+
+### Windows
+
+Support for Windows is currently experimental. Use `nmake` to build:
+
+```cmd
+set "PGROOT=C:\Program Files\PostgreSQL\15"
+git clone --branch v0.4.1 https://github.com/pgvector/pgvector.git
+cd pgvector
+nmake /F Makefile.win
+nmake /F Makefile.win install
+```
 
 ## Additional Installation Methods
 
@@ -227,12 +394,12 @@ Get the [Docker image](https://hub.docker.com/r/ankane/pgvector) with:
 docker pull ankane/pgvector
 ```
 
-This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres).
+This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (run it the same way).
 
-You can also build the image manually
+You can also build the image manually:
 
 ```sh
-git clone --branch v0.3.2 https://github.com/pgvector/pgvector.git
+git clone --branch v0.4.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build -t pgvector .
 ```
@@ -242,7 +409,7 @@ docker build -t pgvector .
 With Homebrew Postgres, you can use:
 
 ```sh
-brew install pgvector/brew/pgvector
+brew install pgvector
 ```
 
 ### PGXN
@@ -253,14 +420,39 @@ Install from the [PostgreSQL Extension Network](https://pgxn.org/dist/vector) wi
 pgxn install vector
 ```
 
+### 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)
+
 ## Hosted Postgres
 
-Some Postgres providers only support specific extensions. To request a new extension:
+pgvector is available on [these providers](https://github.com/pgvector/pgvector/issues/54).
+
+To request a new extension on other providers:
 
 - Amazon RDS - follow the instructions on [this page](https://aws.amazon.com/rds/postgresql/faqs/)
-- Google Cloud SQL - follow the instructions on [this page](https://cloud.google.com/sql/docs/postgres/extensions#requesting-support-for-a-new-extension)
+- Google Cloud SQL - vote or comment on [this page](https://issuetracker.google.com/issues/265172065)
+- Azure Database - vote or comment on [this page](https://feedback.azure.com/d365community/idea/7b423322-6189-ed11-a81b-000d3ae49307)
 - DigitalOcean Managed Databases - vote or comment on [this page](https://ideas.digitalocean.com/app-framework-services/p/pgvector-extension-for-postgresql)
-- Azure Database for PostgreSQL - follow the instructions on [this page](https://docs.microsoft.com/en-us/azure/postgresql/concepts-extensions#next-steps)
+- Heroku Postgres - vote or comment on [this page](https://github.com/heroku/roadmap/issues/156)
 
 ## Upgrading
 
@@ -272,6 +464,16 @@ ALTER EXTENSION vector UPDATE;
 
 ## 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/vector--0.2.0--0.2.1.sql

@@ -7,13 +7,13 @@ 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;
+	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;
+	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;
+	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;
+	WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;

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.sql

@@ -25,7 +25,8 @@ CREATE TYPE vector (
 	OUTPUT    = vector_out,
 	TYPMOD_IN = vector_typmod_in,
 	RECEIVE   = vector_recv,
-	SEND      = vector_send
+	SEND      = vector_send,
+	STORAGE   = extended
 );
 
 -- functions
@@ -83,6 +84,26 @@ CREATE FUNCTION vector_negative_inner_product(vector, vector) RETURNS float8
 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
+);
+
 -- cast functions
 
 CREATE FUNCTION vector(vector, integer, boolean) RETURNS vector

src/ivfbuild.c

@@ -6,6 +6,7 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
+#include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"
@@ -22,13 +23,8 @@
 #define PROGRESS_CREATEIDX_TUPLES_DONE 0
 #endif
 
-#if PG_VERSION_NUM >= 110000
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
-#else
-#include "catalog/pg_operator.h"
-#include "catalog/pg_type.h"
-#endif
 
 #if PG_VERSION_NUM >= 130000
 #define CALLBACK_ITEM_POINTER ItemPointer tid
@@ -43,20 +39,16 @@
 #endif
 
 /*
- * Callback for sampling
+ * Add sample
  */
 static void
-SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
-			   bool *isnull, bool tupleIsAlive, void *state)
+AddSample(Datum *values, IvfflatBuildState * buildstate)
 {
-	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
 	VectorArray samples = buildstate->samples;
 	int			targsamples = samples->maxlen;
-	Datum		value = values[0];
 
-	/* Skip nulls */
-	if (isnull[0])
-		return;
+	/* Detoast once for all calls */
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/*
 	 * Normalize with KMEANS_NORM_PROC since spherical distance function
@@ -94,6 +86,31 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	}
 }
 
+/*
+ * Callback for sampling
+ */
+static void
+SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
+			   bool *isnull, bool tupleIsAlive, void *state)
+{
+	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
+	MemoryContext oldCtx;
+
+	/* Skip nulls */
+	if (isnull[0])
+		return;
+
+	/* Use memory context since detoast can allocate */
+	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
+
+	/* Add sample */
+	AddSample(values, state);
+
+	/* Reset memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextReset(buildstate->tmpCtx);
+}
+
 /*
  * Sample rows with same logic as ANALYZE
  */
@@ -103,11 +120,9 @@ SampleRows(IvfflatBuildState * buildstate)
 	int			targsamples = buildstate->samples->maxlen;
 	BlockNumber totalblocks = RelationGetNumberOfBlocks(buildstate->heap);
 
-	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_SAMPLE);
-
 	buildstate->rowstoskip = -1;
 
-	BlockSampler_Init(&buildstate->bs, totalblocks, targsamples, random());
+	BlockSampler_Init(&buildstate->bs, totalblocks, targsamples, RandomInt());
 
 	reservoir_init_selection_state(&buildstate->rstate, targsamples);
 	while (BlockSampler_HasMore(&buildstate->bs))
@@ -117,38 +132,28 @@ SampleRows(IvfflatBuildState * buildstate)
 #if PG_VERSION_NUM >= 120000
 		table_index_build_range_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
 									 false, true, false, targblock, 1, SampleCallback, (void *) buildstate, NULL);
-#elif PG_VERSION_NUM >= 110000
-		IndexBuildHeapRangeScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-								false, true, targblock, 1, SampleCallback, (void *) buildstate, NULL);
 #else
 		IndexBuildHeapRangeScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-								false, true, targblock, 1, SampleCallback, (void *) buildstate);
+								false, true, targblock, 1, SampleCallback, (void *) buildstate, NULL);
 #endif
 	}
 }
 
 /*
- * Callback for table_index_build_scan
+ * Add tuple to sort
  */
 static void
-BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
-			  bool *isnull, bool tupleIsAlive, void *state)
+AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState * buildstate)
 {
-	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
 	double		distance;
 	double		minDistance = DBL_MAX;
 	int			closestCenter = -1;
 	VectorArray centers = buildstate->centers;
 	TupleTableSlot *slot = buildstate->slot;
-	Datum		value = values[0];
 	int			i;
 
-#if PG_VERSION_NUM < 130000
-	ItemPointer tid = &hup->t_self;
-#endif
-
-	if (isnull[0])
-		return;
+	/* Detoast once for all calls */
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
@@ -196,6 +201,35 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	buildstate->indtuples++;
 }
 
+/*
+ * Callback for table_index_build_scan
+ */
+static void
+BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
+			  bool *isnull, bool tupleIsAlive, void *state)
+{
+	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
+	MemoryContext oldCtx;
+
+#if PG_VERSION_NUM < 130000
+	ItemPointer tid = &hup->t_self;
+#endif
+
+	/* Skip nulls */
+	if (isnull[0])
+		return;
+
+	/* Use memory context since detoast can allocate */
+	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
+
+	/* Add tuple to sort */
+	AddTupleToSort(index, tid, values, buildstate);
+
+	/* Reset memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextReset(buildstate->tmpCtx);
+}
+
 /*
  * Get index tuple from sort state
  */
@@ -205,11 +239,7 @@ GetNextTuple(Tuplesortstate *sortstate, TupleDesc tupdesc, TupleTableSlot *slot,
 	Datum		value;
 	bool		isnull;
 
-#if PG_VERSION_NUM >= 100000
 	if (tuplesort_gettupleslot(sortstate, true, false, slot, NULL))
-#else
-	if (tuplesort_gettupleslot(sortstate, true, slot, NULL))
-#endif
 	{
 		*list = DatumGetInt32(slot_getattr(slot, 1, &isnull));
 		value = slot_getattr(slot, 3, &isnull);
@@ -308,6 +338,9 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	if (buildstate->dimensions < 0)
 		elog(ERROR, "column does not have dimensions");
 
+	if (buildstate->dimensions > IVFFLAT_MAX_DIM)
+		elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", IVFFLAT_MAX_DIM);
+
 	buildstate->reltuples = 0;
 	buildstate->indtuples = 0;
 
@@ -331,11 +364,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 #endif
 	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
 	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
-#if PG_VERSION_NUM >= 110000
 	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 3, "vector", RelationGetDescr(index)->attrs[0].atttypid, -1, 0);
-#else
-	TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 3, "vector", RelationGetDescr(index)->attrs[0]->atttypid, -1, 0);
-#endif
 
 #if PG_VERSION_NUM >= 120000
 	buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsVirtual);
@@ -349,6 +378,10 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	/* Reuse for each tuple */
 	buildstate->normvec = InitVector(buildstate->dimensions);
 
+	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
+											   "Ivfflat build temporary context",
+											   ALLOCSET_DEFAULT_SIZES);
+
 #ifdef IVFFLAT_KMEANS_DEBUG
 	buildstate->inertia = 0;
 	buildstate->listSums = palloc0(sizeof(double) * buildstate->lists);
@@ -362,7 +395,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 static void
 FreeBuildState(IvfflatBuildState * buildstate)
 {
-	pfree(buildstate->centers);
+	VectorArrayFree(buildstate->centers);
 	pfree(buildstate->listInfo);
 	pfree(buildstate->normvec);
 
@@ -370,6 +403,8 @@ FreeBuildState(IvfflatBuildState * buildstate)
 	pfree(buildstate->listSums);
 	pfree(buildstate->listCounts);
 #endif
+
+	MemoryContextDelete(buildstate->tmpCtx);
 }
 
 /*
@@ -380,6 +415,8 @@ ComputeCenters(IvfflatBuildState * buildstate)
 {
 	int			numSamples;
 
+	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
+
 	/* Target 50 samples per list, with at least 10000 samples */
 	/* The number of samples has a large effect on index build time */
 	numSamples = buildstate->lists * 50;
@@ -394,14 +431,23 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	/* TODO Ensure within maintenance_work_mem */
 	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions);
 	if (buildstate->heap != NULL)
+	{
 		SampleRows(buildstate);
 
+		if (buildstate->samples->length < buildstate->lists)
+		{
+			ereport(NOTICE,
+					(errmsg("ivfflat index created with little data"),
+					 errdetail("this will cause poor recall"),
+					 errhint("drop the index until the table has more data")));
+		}
+	}
+
 	/* Calculate centers */
-	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
 	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers));
 
 	/* Free samples before we allocate more memory */
-	pfree(buildstate->samples);
+	VectorArrayFree(buildstate->samples);
 }
 
 /*
@@ -535,11 +581,7 @@ CreateEntryPages(IvfflatBuildState * buildstate, ForkNumber forkNum)
 
 	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_SORT);
 
-#if PG_VERSION_NUM >= 110000
 	buildstate->sortstate = tuplesort_begin_heap(buildstate->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, maintenance_work_mem, NULL, false);
-#else
-	buildstate->sortstate = tuplesort_begin_heap(buildstate->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, maintenance_work_mem, false);
-#endif
 
 	/* Add tuples to sort */
 	if (buildstate->heap != NULL)
@@ -547,12 +589,9 @@ CreateEntryPages(IvfflatBuildState * buildstate, ForkNumber forkNum)
 #if PG_VERSION_NUM >= 120000
 		buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
 													   true, true, BuildCallback, (void *) buildstate, NULL);
-#elif PG_VERSION_NUM >= 110000
-		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, BuildCallback, (void *) buildstate, NULL);
 #else
 		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, BuildCallback, (void *) buildstate);
+												   true, BuildCallback, (void *) buildstate, NULL);
 #endif
 	}
 

src/ivfflat.c

@@ -45,8 +45,6 @@ ivfflatbuildphasename(int64 phasenum)
 	{
 		case PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE:
 			return "initializing";
-		case PROGRESS_IVFFLAT_PHASE_SAMPLE:
-			return "sampling table";
 		case PROGRESS_IVFFLAT_PHASE_KMEANS:
 			return "performing k-means";
 		case PROGRESS_IVFFLAT_PHASE_SORT:
@@ -66,9 +64,7 @@ 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
 )
 {
 	GenericCosts costs;
@@ -86,9 +82,7 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 		*indexTotalCost = DBL_MAX;
 		*indexSelectivity = 0;
 		*indexCorrelation = 0;
-#if PG_VERSION_NUM >= 100000
 		*indexPages = 0;
-#endif
 		return;
 	}
 
@@ -116,9 +110,7 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	*indexTotalCost = costs.indexTotalCost;
 	*indexSelectivity = costs.indexSelectivity;
 	*indexCorrelation = costs.indexCorrelation;
-#if PG_VERSION_NUM >= 100000
 	*indexPages = costs.numIndexPages;
-#endif
 }
 
 /*
@@ -164,7 +156,7 @@ ivfflatvalidate(Oid opclassoid)
  *
  * See https://www.postgresql.org/docs/current/index-api.html
  */
-PG_FUNCTION_INFO_V1(ivfflathandler);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(ivfflathandler);
 Datum
 ivfflathandler(PG_FUNCTION_ARGS)
 {
@@ -186,12 +178,8 @@ 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_PARALLEL_BULKDEL;
@@ -224,11 +212,9 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	amroutine->amrestrpos = NULL;
 
 	/* Interface functions to support parallel index scans */
-#if PG_VERSION_NUM >= 100000
 	amroutine->amestimateparallelscan = NULL;
 	amroutine->aminitparallelscan = NULL;
 	amroutine->amparallelrescan = NULL;
-#endif
 
 	PG_RETURN_POINTER(amroutine);
 }

src/ivfflat.h

@@ -3,20 +3,27 @@
 
 #include "postgres.h"
 
+#if PG_VERSION_NUM < 110000
+#error "Requires PostgreSQL 11+"
+#endif
+
 #include "access/generic_xlog.h"
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
+#include "port.h"				/* for strtof() and random() */
 #include "utils/sampling.h"
 #include "utils/tuplesort.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 150000
+#include "common/pg_prng.h"
+#endif
+
 #ifdef IVFFLAT_BENCH
 #include "portability/instr_time.h"
 #endif
 
-#if PG_VERSION_NUM < 100000
-#error "Requires PostgreSQL 10+"
-#endif
+#define IVFFLAT_MAX_DIM 2000
 
 /* Support functions */
 #define IVFFLAT_DISTANCE_PROC 1
@@ -37,10 +44,9 @@
 
 /* Build phases */
 /* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
-#define PROGRESS_IVFFLAT_PHASE_SAMPLE	2
-#define PROGRESS_IVFFLAT_PHASE_KMEANS	3
-#define PROGRESS_IVFFLAT_PHASE_SORT		4
-#define PROGRESS_IVFFLAT_PHASE_LOAD		5
+#define PROGRESS_IVFFLAT_PHASE_KMEANS	2
+#define PROGRESS_IVFFLAT_PHASE_SORT		3
+#define PROGRESS_IVFFLAT_PHASE_LOAD		4
 
 #define IVFFLAT_LIST_SIZE(_dim)	(offsetof(IvfflatListData, center) + VECTOR_SIZE(_dim))
 
@@ -62,15 +68,26 @@
 #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 */
 extern int	ivfflat_probes;
 
+/* Exported functions */
+PGDLLEXPORT void _PG_init(void);
+
 typedef struct VectorArrayData
 {
 	int			length;
 	int			maxlen;
 	int			dim;
-	Vector		items[FLEXIBLE_ARRAY_MEMBER];
+	Vector	   *items;
 }			VectorArrayData;
 
 typedef VectorArrayData * VectorArray;
@@ -130,6 +147,9 @@ typedef struct IvfflatBuildState
 	Tuplesortstate *sortstate;
 	TupleDesc	tupdesc;
 	TupleTableSlot *slot;
+
+	/* Memory */
+	MemoryContext tmpCtx;
 }			IvfflatBuildState;
 
 typedef struct IvfflatMetaPageData
@@ -186,19 +206,20 @@ typedef struct IvfflatScanOpaqueData
 
 	/* Lists */
 	pairingheap *listQueue;
+	double		minDistance;
 	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);
@@ -218,9 +239,7 @@ bool		ivfflatinsert(Relation index, Datum *values, bool *isnull, ItemPointer hea
 #if PG_VERSION_NUM >= 140000
 						  ,bool indexUnchanged
 #endif
-#if PG_VERSION_NUM >= 100000
 						  ,IndexInfo *indexInfo
-#endif
 );
 IndexBulkDeleteResult *ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats, IndexBulkDeleteCallback callback, void *callback_state);
 IndexBulkDeleteResult *ivfflatvacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats);

src/ivfinsert.c

@@ -4,6 +4,7 @@
 
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "utils/memutils.h"
 
 /*
  * Find the list that minimizes the distance function
@@ -57,8 +58,11 @@ FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo *
  * 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;
@@ -67,11 +71,27 @@ InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
 	ListInfo	listInfo;
 	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)));
 
@@ -130,9 +150,14 @@ InsertTuple(Relation rel, IndexTuple itup, Relation heapRel, Datum *values)
 			/* Unlock extend relation lock as early as possible */
 			UnlockReleaseBuffer(metabuf);
 
-			/* Unlock rest */
-			UnlockReleaseBuffer(newbuf);
+			/* Unlock previous buffer */
 			UnlockReleaseBuffer(buf);
+
+			/* Prepare new buffer */
+			state = GenericXLogStart(rel);
+			buf = newbuf;
+			page = GenericXLogRegisterBuffer(state, buf, 0);
+			break;
 		}
 	}
 
@@ -156,36 +181,31 @@ ivfflatinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 #if PG_VERSION_NUM >= 140000
 			  ,bool indexUnchanged
 #endif
-#if PG_VERSION_NUM >= 100000
 			  ,IndexInfo *indexInfo
-#endif
 )
 {
-	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

@@ -5,16 +5,6 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 
-#if PG_VERSION_NUM >= 150000
-#include "common/pg_prng.h"
-#endif
-
-#if PG_VERSION_NUM >= 150000
-#define RandomDouble() pg_prng_double(&pg_global_prng_state)
-#else
-#define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
-#endif
-
 /*
  * Initialize with kmeans++
  *
@@ -39,7 +29,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 	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++)
@@ -453,7 +443,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 			break;
 	}
 
-	pfree(newCenters);
+	VectorArrayFree(newCenters);
 	pfree(centerCounts);
 	pfree(closestCenters);
 	pfree(lowerBound);

src/ivfscan.c

@@ -5,15 +5,11 @@
 #include "access/relscan.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
+#include "pgstat.h"
 #include "storage/bufmgr.h"
 
-#if PG_VERSION_NUM >= 110000
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
-#else
-#include "catalog/pg_operator.h"
-#include "catalog/pg_type.h"
-#endif
 
 /*
  * Compare list distances
@@ -64,6 +60,9 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			/* Use procinfo from the index instead of scan key for performance */
 			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
+			if (distance < so->minDistance)
+				so->minDistance = distance;
+
 			if (listCount < so->probes)
 			{
 				scanlist = &so->lists[listCount];
@@ -115,6 +114,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	Datum		datum;
 	bool		isnull;
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
+	double		tuples = 0;
 
 #if PG_VERSION_NUM >= 120000
 	TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
@@ -132,7 +132,13 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	/* Search closest probes lists */
 	while (!pairingheap_is_empty(so->listQueue))
 	{
-		searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
+		IvfflatScanList *scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
+
+		/* Query-aware dynamic pruning */
+		if (scanlist->distance > 2 * so->minDistance)
+			continue;
+
+		searchPage = scanlist->startPage;
 
 		/* Search all entry pages for list */
 		while (BlockNumberIsValid(searchPage))
@@ -163,6 +169,8 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
+
+				tuples++;
 			}
 
 			searchPage = IvfflatPageGetOpaque(page)->nextblkno;
@@ -171,6 +179,13 @@ GetScanItems(IndexScanDesc scan, Datum value)
 		}
 	}
 
+	/* TODO Scan more lists */
+	if (tuples < 100)
+		ereport(DEBUG1,
+				(errmsg("index scan found few tuples"),
+				 errdetail("index may have been created without data or lists is too high"),
+				 errhint("recreate the index and possibly decrease lists")));
+
 	tuplesort_performsort(so->sortstate);
 }
 
@@ -216,11 +231,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	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);
@@ -250,6 +261,7 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 
 	so->first = true;
 	pairingheap_reset(so->listQueue);
+	so->minDistance = DBL_MAX;
 
 	if (keys && scan->numberOfKeys > 0)
 		memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));
@@ -276,6 +288,9 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 	{
 		Datum		value;
 
+		/* Count index scan for stats */
+		pgstat_count_index_scan(scan->indexRelation);
+
 		/* Safety check */
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
@@ -286,6 +301,10 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 
 		value = scan->orderByData->sk_argument;
 
+		/* Value should not be compressed or toasted */
+		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
+
 		if (so->normprocinfo != NULL)
 		{
 			/* No items will match if normalization fails */
@@ -302,11 +321,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 			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
 	{
 		ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
 		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));

src/ivfutils.c

@@ -10,14 +10,25 @@
 VectorArray
 VectorArrayInit(int maxlen, int dimensions)
 {
-	VectorArray res = palloc_extended(VECTOR_ARRAY_SIZE(maxlen, dimensions), MCXT_ALLOC_ZERO | MCXT_ALLOC_HUGE);
+	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
  */
@@ -75,7 +86,7 @@ IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * resul
 
 	if (norm > 0)
 	{
-		v = (Vector *) DatumGetPointer(*value);
+		v = DatumGetVector(*value);
 
 		if (result == NULL)
 			result = InitVector(v->dim);

src/ivfvacuum.c

@@ -143,6 +143,11 @@ 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.c

@@ -13,13 +13,20 @@
 #include "utils/numeric.h"
 
 #if PG_VERSION_NUM >= 120000
+#include "common/shortest_dec.h"
 #include "utils/float.h"
+#else
+#include <float.h>
 #endif
 
 #if PG_VERSION_NUM < 130000
+#define TYPALIGN_DOUBLE 'd'
 #define TYPALIGN_INT 'i'
 #endif
 
+#define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
+#define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
+
 PG_MODULE_MAGIC;
 
 /*
@@ -79,6 +86,30 @@ CheckElement(float value)
 				 errmsg("infinite value not allowed in vector")));
 }
 
+/*
+ * Check state array
+ */
+static float8 *
+CheckStateArray(ArrayType *statearray, const char *caller)
+{
+	if (ARR_NDIM(statearray) != 1 ||
+		ARR_DIMS(statearray)[0] < 1 ||
+		ARR_HASNULL(statearray) ||
+		ARR_ELEMTYPE(statearray) != FLOAT8OID)
+		elog(ERROR, "%s: expected state array", caller);
+	return (float8 *) ARR_DATA_PTR(statearray);
+}
+
+#if PG_VERSION_NUM < 120003
+static pg_noinline void
+float_overflow_error(void)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+			 errmsg("value out of range: overflow")));
+}
+#endif
+
 /*
  * Print vector - useful for debugging
  */
@@ -106,14 +137,14 @@ PrintVector(char *msg, Vector * vector)
 /*
  * Convert textual representation to internal representation
  */
-PG_FUNCTION_INFO_V1(vector_in);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_in);
 Datum
 vector_in(PG_FUNCTION_ARGS)
 {
 	char	   *str = PG_GETARG_CSTRING(0);
 	int32		typmod = PG_GETARG_INT32(2);
 	int			i;
-	double		x[VECTOR_MAX_DIM];
+	float		x[VECTOR_MAX_DIM];
 	int			dim = 0;
 	char	   *pt;
 	char	   *stringEnd;
@@ -136,7 +167,8 @@ vector_in(PG_FUNCTION_ARGS)
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 					 errmsg("vector cannot have more than %d dimensions", VECTOR_MAX_DIM)));
 
-		x[dim] = strtod(pt, &stringEnd);
+		/* Use strtof like float4in to avoid a double-rounding problem */
+		x[dim] = strtof(pt, &stringEnd);
 		CheckElement(x[dim]);
 		dim++;
 
@@ -182,35 +214,68 @@ vector_in(PG_FUNCTION_ARGS)
 /*
  * Convert internal representation to textual representation
  */
-PG_FUNCTION_INFO_V1(vector_out);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_out);
 Datum
 vector_out(PG_FUNCTION_ARGS)
 {
 	Vector	   *vector = PG_GETARG_VECTOR_P(0);
-	StringInfoData buf;
 	int			dim = vector->dim;
+	char	   *buf;
+	char	   *ptr;
 	int			i;
+	int			n;
 
-	initStringInfo(&buf);
+#if PG_VERSION_NUM < 120000
+	int			ndig = FLT_DIG + extra_float_digits;
 
-	appendStringInfoChar(&buf, '[');
+	if (ndig < 1)
+		ndig = 1;
+
+#define FLOAT_SHORTEST_DECIMAL_LEN (ndig + 10)
+#endif
+
+	/*
+	 * Need:
+	 *
+	 * dim * (FLOAT_SHORTEST_DECIMAL_LEN - 1) bytes for
+	 * float_to_shortest_decimal_bufn
+	 *
+	 * dim - 1 bytes for separator
+	 *
+	 * 3 bytes for [, ], and \0
+	 */
+	buf = (char *) palloc(FLOAT_SHORTEST_DECIMAL_LEN * dim + 2);
+	ptr = buf;
+
+	*ptr = '[';
+	ptr++;
 	for (i = 0; i < dim; i++)
 	{
 		if (i > 0)
-			appendStringInfoString(&buf, ",");
+		{
+			*ptr = ',';
+			ptr++;
+		}
 
-		appendStringInfoString(&buf, float8out_internal(vector->x[i]));
+#if PG_VERSION_NUM >= 120000
+		n = float_to_shortest_decimal_bufn(vector->x[i], ptr);
+#else
+		n = sprintf(ptr, "%.*g", ndig, vector->x[i]);
+#endif
+		ptr += n;
 	}
-	appendStringInfoChar(&buf, ']');
+	*ptr = ']';
+	ptr++;
+	*ptr = '\0';
 
 	PG_FREE_IF_COPY(vector, 0);
-	PG_RETURN_CSTRING(buf.data);
+	PG_RETURN_CSTRING(buf);
 }
 
 /*
  * Convert type modifier
  */
-PG_FUNCTION_INFO_V1(vector_typmod_in);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_typmod_in);
 Datum
 vector_typmod_in(PG_FUNCTION_ARGS)
 {
@@ -241,7 +306,7 @@ vector_typmod_in(PG_FUNCTION_ARGS)
 /*
  * Convert external binary representation to internal representation
  */
-PG_FUNCTION_INFO_V1(vector_recv);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_recv);
 Datum
 vector_recv(PG_FUNCTION_ARGS)
 {
@@ -273,7 +338,7 @@ vector_recv(PG_FUNCTION_ARGS)
 /*
  * Convert internal representation to the external binary representation
  */
-PG_FUNCTION_INFO_V1(vector_send);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_send);
 Datum
 vector_send(PG_FUNCTION_ARGS)
 {
@@ -293,7 +358,7 @@ vector_send(PG_FUNCTION_ARGS)
 /*
  * Convert vector to vector
  */
-PG_FUNCTION_INFO_V1(vector);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector);
 Datum
 vector(PG_FUNCTION_ARGS)
 {
@@ -308,7 +373,7 @@ vector(PG_FUNCTION_ARGS)
 /*
  * Convert array to vector
  */
-PG_FUNCTION_INFO_V1(array_to_vector);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(array_to_vector);
 Datum
 array_to_vector(PG_FUNCTION_ARGS)
 {
@@ -351,7 +416,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 		else if (ARR_ELEMTYPE(array) == FLOAT4OID)
 			result->x[i] = DatumGetFloat4(elemsp[i]);
 		else if (ARR_ELEMTYPE(array) == NUMERICOID)
-			result->x[i] = DatumGetFloat4(DirectFunctionCall1(numeric_float4, NumericGetDatum(elemsp[i])));
+			result->x[i] = DatumGetFloat4(DirectFunctionCall1(numeric_float4, elemsp[i]));
 		else
 			ereport(ERROR,
 					(errcode(ERRCODE_DATA_EXCEPTION),
@@ -366,7 +431,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 /*
  * Convert vector to float4[]
  */
-PG_FUNCTION_INFO_V1(vector_to_float4);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_to_float4);
 Datum
 vector_to_float4(PG_FUNCTION_ARGS)
 {
@@ -389,20 +454,23 @@ vector_to_float4(PG_FUNCTION_ARGS)
 /*
  * Get the L2 distance between vectors
  */
-PG_FUNCTION_INFO_V1(l2_distance);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(l2_distance);
 Datum
 l2_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	double		distance = 0.0;
 	double		diff;
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
 	{
-		diff = a->x[i] - b->x[i];
+		diff = ax[i] - bx[i];
 		distance += diff * diff;
 	}
 
@@ -413,20 +481,23 @@ l2_distance(PG_FUNCTION_ARGS)
  * Get the L2 squared distance between vectors
  * This saves a sqrt calculation
  */
-PG_FUNCTION_INFO_V1(vector_l2_squared_distance);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_l2_squared_distance);
 Datum
 vector_l2_squared_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	double		distance = 0.0;
 	double		diff;
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
 	{
-		diff = a->x[i] - b->x[i];
+		diff = ax[i] - bx[i];
 		distance += diff * diff;
 	}
 
@@ -436,18 +507,21 @@ vector_l2_squared_distance(PG_FUNCTION_ARGS)
 /*
  * Get the inner product of two vectors
  */
-PG_FUNCTION_INFO_V1(inner_product);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(inner_product);
 Datum
 inner_product(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	double		distance = 0.0;
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
-		distance += a->x[i] * b->x[i];
+		distance += ax[i] * bx[i];
 
 	PG_RETURN_FLOAT8(distance);
 }
@@ -455,18 +529,21 @@ inner_product(PG_FUNCTION_ARGS)
 /*
  * Get the negative inner product of two vectors
  */
-PG_FUNCTION_INFO_V1(vector_negative_inner_product);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_negative_inner_product);
 Datum
 vector_negative_inner_product(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	double		distance = 0.0;
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
-		distance += a->x[i] * b->x[i];
+		distance += ax[i] * bx[i];
 
 	PG_RETURN_FLOAT8(distance * -1);
 }
@@ -474,26 +551,30 @@ vector_negative_inner_product(PG_FUNCTION_ARGS)
 /*
  * Get the cosine distance between two vectors
  */
-PG_FUNCTION_INFO_V1(cosine_distance);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(cosine_distance);
 Datum
 cosine_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	double		distance = 0.0;
 	double		norma = 0.0;
 	double		normb = 0.0;
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
 	{
-		distance += a->x[i] * b->x[i];
-		norma += a->x[i] * a->x[i];
-		normb += b->x[i] * b->x[i];
+		distance += ax[i] * bx[i];
+		norma += ax[i] * ax[i];
+		normb += bx[i] * bx[i];
 	}
 
-	PG_RETURN_FLOAT8(1 - (distance / (sqrt(norma) * sqrt(normb))));
+	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
+	PG_RETURN_FLOAT8(1 - (distance / sqrt(norma * normb)));
 }
 
 /*
@@ -501,7 +582,7 @@ cosine_distance(PG_FUNCTION_ARGS)
  * Currently uses angular distance since needs to satisfy triangle inequality
  * Assumes inputs are unit vectors (skips norm)
  */
-PG_FUNCTION_INFO_V1(vector_spherical_distance);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_spherical_distance);
 Datum
 vector_spherical_distance(PG_FUNCTION_ARGS)
 {
@@ -511,6 +592,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
 		distance += a->x[i] * b->x[i];
 
@@ -526,7 +608,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 /*
  * Get the dimensions of a vector
  */
-PG_FUNCTION_INFO_V1(vector_dims);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_dims);
 Datum
 vector_dims(PG_FUNCTION_ARGS)
 {
@@ -538,15 +620,17 @@ vector_dims(PG_FUNCTION_ARGS)
 /*
  * Get the L2 norm of a vector
  */
-PG_FUNCTION_INFO_V1(vector_norm);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_norm);
 Datum
 vector_norm(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	float	   *ax = a->x;
 	double		norm = 0.0;
 
+	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
-		norm += a->x[i] * a->x[i];
+		norm += ax[i] * ax[i];
 
 	PG_RETURN_FLOAT8(sqrt(norm));
 }
@@ -554,20 +638,25 @@ vector_norm(PG_FUNCTION_ARGS)
 /*
  * Add vectors
  */
-PG_FUNCTION_INFO_V1(vector_add);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_add);
 Datum
 vector_add(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	Vector	   *result;
-	int			i;
+	float	   *rx;
 
 	CheckDims(a, b);
 
 	result = InitVector(a->dim);
-	for (i = 0; i < a->dim; i++)
-		result->x[i] = a->x[i] + b->x[i];
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+		rx[i] = ax[i] + bx[i];
 
 	PG_RETURN_POINTER(result);
 }
@@ -575,20 +664,25 @@ vector_add(PG_FUNCTION_ARGS)
 /*
  * Subtract vectors
  */
-PG_FUNCTION_INFO_V1(vector_sub);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_sub);
 Datum
 vector_sub(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
 	Vector	   *result;
-	int			i;
+	float	   *rx;
 
 	CheckDims(a, b);
 
 	result = InitVector(a->dim);
-	for (i = 0; i < a->dim; i++)
-		result->x[i] = a->x[i] - b->x[i];
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+		rx[i] = ax[i] - bx[i];
 
 	PG_RETURN_POINTER(result);
 }
@@ -617,7 +711,7 @@ vector_cmp_internal(Vector * a, Vector * b)
 /*
  * Less than
  */
-PG_FUNCTION_INFO_V1(vector_lt);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_lt);
 Datum
 vector_lt(PG_FUNCTION_ARGS)
 {
@@ -630,7 +724,7 @@ vector_lt(PG_FUNCTION_ARGS)
 /*
  * Less than or equal
  */
-PG_FUNCTION_INFO_V1(vector_le);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_le);
 Datum
 vector_le(PG_FUNCTION_ARGS)
 {
@@ -643,7 +737,7 @@ vector_le(PG_FUNCTION_ARGS)
 /*
  * Equal
  */
-PG_FUNCTION_INFO_V1(vector_eq);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_eq);
 Datum
 vector_eq(PG_FUNCTION_ARGS)
 {
@@ -656,7 +750,7 @@ vector_eq(PG_FUNCTION_ARGS)
 /*
  * Not equal
  */
-PG_FUNCTION_INFO_V1(vector_ne);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ne);
 Datum
 vector_ne(PG_FUNCTION_ARGS)
 {
@@ -669,7 +763,7 @@ vector_ne(PG_FUNCTION_ARGS)
 /*
  * Greater than or equal
  */
-PG_FUNCTION_INFO_V1(vector_ge);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ge);
 Datum
 vector_ge(PG_FUNCTION_ARGS)
 {
@@ -682,7 +776,7 @@ vector_ge(PG_FUNCTION_ARGS)
 /*
  * Greater than
  */
-PG_FUNCTION_INFO_V1(vector_gt);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_gt);
 Datum
 vector_gt(PG_FUNCTION_ARGS)
 {
@@ -695,7 +789,7 @@ vector_gt(PG_FUNCTION_ARGS)
 /*
  * Compare vectors
  */
-PG_FUNCTION_INFO_V1(vector_cmp);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_cmp);
 Datum
 vector_cmp(PG_FUNCTION_ARGS)
 {
@@ -704,3 +798,167 @@ vector_cmp(PG_FUNCTION_ARGS)
 
 	PG_RETURN_INT32(vector_cmp_internal(a, b));
 }
+
+/*
+ * Accumulate vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_accum);
+Datum
+vector_accum(PG_FUNCTION_ARGS)
+{
+	ArrayType  *statearray = PG_GETARG_ARRAYTYPE_P(0);
+	Vector	   *newval = PG_GETARG_VECTOR_P(1);
+	float8	   *statevalues;
+	int16		dim;
+	bool		newarr;
+	float8		n;
+	Datum	   *statedatums;
+	float	   *x = newval->x;
+	ArrayType  *result;
+
+	/* Check array before using */
+	statevalues = CheckStateArray(statearray, "vector_accum");
+	dim = STATE_DIMS(statearray);
+	newarr = dim == 0;
+
+	if (newarr)
+		dim = newval->dim;
+	else
+		CheckExpectedDim(dim, newval->dim);
+
+	n = statevalues[0] + 1.0;
+
+	statedatums = CreateStateDatums(dim);
+	statedatums[0] = Float8GetDatumFast(n);
+
+	if (newarr)
+	{
+		for (int i = 0; i < dim; i++)
+			statedatums[i + 1] = Float8GetDatumFast((double) x[i]);
+	}
+	else
+	{
+		for (int i = 0; i < dim; i++)
+		{
+			double		v = statevalues[i + 1] + x[i];
+
+			if (isinf(v))
+				float_overflow_error();
+
+			statedatums[i + 1] = Float8GetDatumFast(v);
+		}
+	}
+
+	/* Use float8 array like float4_accum */
+	result = construct_array(statedatums, dim + 1,
+							 FLOAT8OID,
+							 sizeof(float8), FLOAT8PASSBYVAL, TYPALIGN_DOUBLE);
+
+	pfree(statedatums);
+
+	PG_RETURN_ARRAYTYPE_P(result);
+}
+
+/*
+ * Combine vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_combine);
+Datum
+vector_combine(PG_FUNCTION_ARGS)
+{
+	ArrayType  *statearray1 = PG_GETARG_ARRAYTYPE_P(0);
+	ArrayType  *statearray2 = PG_GETARG_ARRAYTYPE_P(1);
+	float8	   *statevalues1;
+	float8	   *statevalues2;
+	float8		n;
+	float8		n1;
+	float8		n2;
+	int16		dim;
+	Datum	   *statedatums;
+	ArrayType  *result;
+
+	/* Check arrays before using */
+	statevalues1 = CheckStateArray(statearray1, "vector_combine");
+	statevalues2 = CheckStateArray(statearray2, "vector_combine");
+
+	n1 = statevalues1[0];
+	n2 = statevalues2[0];
+
+	if (n1 == 0.0)
+	{
+		n = n2;
+		dim = STATE_DIMS(statearray2);
+		statedatums = CreateStateDatums(dim);
+		for (int i = 1; i <= dim; i++)
+			statedatums[i] = Float8GetDatumFast(statevalues2[i]);
+	}
+	else if (n2 == 0.0)
+	{
+		n = n1;
+		dim = STATE_DIMS(statearray1);
+		statedatums = CreateStateDatums(dim);
+		for (int i = 1; i <= dim; i++)
+			statedatums[i] = Float8GetDatumFast(statevalues1[i]);
+	}
+	else
+	{
+		n = n1 + n2;
+		dim = STATE_DIMS(statearray1);
+		CheckExpectedDim(dim, STATE_DIMS(statearray2));
+		statedatums = CreateStateDatums(dim);
+		for (int i = 1; i <= dim; i++)
+		{
+			double		v = statevalues1[i] + statevalues2[i];
+
+			if (isinf(v))
+				float_overflow_error();
+
+			statedatums[i] = Float8GetDatumFast(v);
+		}
+	}
+
+	statedatums[0] = Float8GetDatumFast(n);
+
+	result = construct_array(statedatums, dim + 1,
+							 FLOAT8OID,
+							 sizeof(float8), FLOAT8PASSBYVAL, TYPALIGN_DOUBLE);
+
+	pfree(statedatums);
+
+	PG_RETURN_ARRAYTYPE_P(result);
+}
+
+/*
+ * Average vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_avg);
+Datum
+vector_avg(PG_FUNCTION_ARGS)
+{
+	ArrayType  *statearray = PG_GETARG_ARRAYTYPE_P(0);
+	float8	   *statevalues;
+	float8		n;
+	uint16		dim;
+	Vector	   *result;
+	float		v;
+
+	/* Check array before using */
+	statevalues = CheckStateArray(statearray, "vector_avg");
+	n = statevalues[0];
+
+	/* SQL defines AVG of no values to be NULL */
+	if (n == 0.0)
+		PG_RETURN_NULL();
+
+	/* Create vector */
+	dim = STATE_DIMS(statearray);
+	result = InitVector(dim);
+	for (int i = 0; i < dim; i++)
+	{
+		v = statevalues[i + 1] / n;
+		CheckElement(v);
+		result->x[i] = v;
+	}
+
+	PG_RETURN_POINTER(result);
+}

src/vector.h

@@ -3,7 +3,11 @@
 
 #include "postgres.h"
 
-#define VECTOR_MAX_DIM 1024
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
+#define VECTOR_MAX_DIM 16000
 
 #define VECTOR_SIZE(_dim)		(offsetof(Vector, x) + sizeof(float)*(_dim))
 #define DatumGetVector(x)		((Vector *) PG_DETOAST_DATUM(x))

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

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SELECT ARRAY[1,2,3]::vector;
   array  
 ---------
@@ -24,6 +22,12 @@ SELECT ARRAY[1,2,3]::float8[]::vector;
  [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 containing NULLs
 SELECT '{NaN}'::real[]::vector;
@@ -40,8 +44,8 @@ SELECT '[1,2,3]'::vector::real[];
  {1,2,3}
 (1 row)
 
-SELECT array_agg(n)::vector FROM generate_series(1, 1025) n;
-ERROR:  vector cannot have more than 1024 dimensions
+SELECT array_agg(n)::vector 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? 

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

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SELECT '[1,2,3]'::vector + '[4,5,6]';
  ?column? 
 ----------
@@ -24,10 +22,28 @@ SELECT round(vector_norm('[1,1]')::numeric, 5);
  1.41421
 (1 row)
 
-SELECT round(l2_distance('[1,2]', '[0,0]')::numeric, 5);
-  round  
----------
- 2.23607
+SELECT vector_norm('[3,4]');
+ vector_norm 
+-------------
+           5
+(1 row)
+
+SELECT vector_norm('[0,1]');
+ vector_norm 
+-------------
+           1
+(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]');
@@ -40,10 +56,10 @@ SELECT inner_product('[1,2]', '[3,4]');
 
 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
+SELECT cosine_distance('[1,2]', '[2,4]');
+ cosine_distance 
+-----------------
+               0
 (1 row)
 
 SELECT cosine_distance('[1,2]', '[0,0]');
@@ -52,5 +68,37 @@ SELECT cosine_distance('[1,2]', '[0,0]');
              NaN
 (1 row)
 
+SELECT cosine_distance('[1,1]', '[1,1]');
+ cosine_distance 
+-----------------
+               0
+(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 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

test/expected/input.out

@@ -1,5 +1,3 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
 SELECT '[1,2,3]'::vector;
  vector  
 ---------
@@ -12,6 +10,12 @@ SELECT '[-1,2,3]'::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"
 LINE 1: SELECT '[hello,1]'::vector;

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);

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);

test/expected/ivfflat_l2.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/ivfflat_options.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));
 CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);

test/expected/ivfflat_unlogged.out

@@ -1,5 +1,3 @@
-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);

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

@@ -1,17 +1,15 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS vector;
-
 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 '[1,2,3]'::vector::real[];
-SELECT array_agg(n)::vector FROM generate_series(1, 1025) n;
+SELECT array_agg(n)::vector 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

@@ -1,18 +1,26 @@
-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 round(vector_norm('[1,1]')::numeric, 5);
+SELECT vector_norm('[3,4]');
+SELECT vector_norm('[0,1]');
+
+SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]', '[0,1]');
 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]', '[2,4]');
 SELECT cosine_distance('[1,2]', '[0,0]');
+SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
 SELECT cosine_distance('[1,2]', '[3]');
+
+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;

test/sql/input.sql

@@ -1,8 +1,6 @@
-SET client_min_messages = warning;
-CREATE EXTENSION IF NOT EXISTS 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;

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));

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));

test/sql/ivfflat_l2.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/ivfflat_options.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/ivfflat_unlogged.sql

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

test/t/001_wal.pl

@@ -20,15 +20,8 @@ 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";
 
@@ -63,6 +56,9 @@ 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';
 

test/t/002_vacuum.pl

@@ -4,6 +4,15 @@ use PostgresNode;
 use TestLib;
 use Test::More tests => 1;
 
+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;
@@ -11,9 +20,9 @@ $node->start;
 
 # Create table and index
 $node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i % 10, ARRAY[i % 1000, i % 333, i % 55] FROM generate_series(1, 100000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
 $node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
 
@@ -24,7 +33,7 @@ my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_id
 $node->safe_psql("postgres", "DELETE FROM tst;");
 $node->safe_psql("postgres", "VACUUM tst;");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i % 10, ARRAY[i % 1000, i % 333, i % 55] FROM generate_series(1, 100000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
 
 # Check size

test/t/007_inserts.pl

@@ -2,7 +2,7 @@ use strict;
 use warnings;
 use PostgresNode;
 use TestLib;
-use Test::More tests => 5;
+use Test::More tests => 7;
 
 my $dim = 768;
 
@@ -32,10 +32,19 @@ $node->pgbench(
 	}
 );
 
+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;
@@ -43,3 +52,4 @@ $count = $node->safe_psql("postgres", qq(
 	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
 ));
 is($count, $expected);
+is(idx_scan(), 1);

test/t/008_avg.pl

@@ -0,0 +1,35 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More tests => 5;
+
+# 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;
+));
+
+# Test avg
+my $avg = $node->safe_psql("postgres", "SELECT AVG(v) FROM tst;");
+like($avg, qr/\[1\.5/);
+like($avg, qr/,2\.5/);
+like($avg, qr/,3\.5/);
+
+# Test matches real
+my $r1 = $node->safe_psql("postgres", "SELECT AVG(r1)::float4 FROM tst;");
+my $r2 = $node->safe_psql("postgres", "SELECT AVG(r2)::float4 FROM tst;");
+my $r3 = $node->safe_psql("postgres", "SELECT AVG(r3)::float4 FROM tst;");
+is($avg, "[$r1,$r2,$r3]");
+
+# Test explain
+my $explain = $node->safe_psql("postgres", "EXPLAIN SELECT AVG(v) FROM tst;");
+like($explain, qr/Partial Aggregate/);

test/t/009_storage.pl

@@ -0,0 +1,32 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More tests => 1;
+
+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/);

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat access method'
-default_version = '0.3.2'
+default_version = '0.4.1'
 module_pathname = '$libdir/vector'
 relocatable = true