Use normalize_l2 for ivfflat

.github/workflows/build.yml

@@ -8,8 +8,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - postgres: 17
-            os: ubuntu-22.04
           - postgres: 16
             os: ubuntu-22.04
           - postgres: 15
@@ -23,7 +21,7 @@ jobs:
           - postgres: 11
             os: ubuntu-20.04
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -45,7 +43,7 @@ jobs:
     runs-on: macos-latest
     if: ${{ !startsWith(github.ref_name, 'windows') }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: 14
@@ -67,7 +65,7 @@ jobs:
     runs-on: windows-latest
     if: ${{ !startsWith(github.ref_name, 'mac') }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: 14
@@ -80,7 +78,6 @@ jobs:
           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

CHANGELOG.md

@@ -1,27 +1,14 @@
-## 0.5.2 (unreleased)
+## 0.5.0 (unreleased)
-
-- Improved performance of HNSW
-- Added support for on-disk parallel index builds for HNSW
-- Reduced memory usage for HNSW index builds
-- Reduced WAL generation for HNSW index builds
-- Fixed error with logical replication
-- Fixed `invalid memory alloc request size` error with HNSW index build
-
-## 0.5.1 (2023-10-10)
-
-- Improved performance of HNSW index builds
-- Added check for MVCC-compliant snapshot for index scans
-
-## 0.5.0 (2023-08-28)
 
 - Added HNSW index type
-- Added support for parallel index builds for IVFFlat
+- Added support for parallel index builds
 - Added `l1_distance` function
+- Added `normalize_l2` 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
+- Fixed results for NULL and NaN distances
 
 ## 0.4.4 (2023-06-12)
 

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": "0.5.1",
+	"version": "0.4.4",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.5.1",
+			"version": "0.4.4",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,14 +1,13 @@
 EXTENSION = vector
-EXTVERSION = 0.5.1
+EXTVERSION = 0.4.4
 
 MODULE_big = vector
 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 test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
-REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
+REGRESS_OPTS = --inputdir=test --load-extension=vector
 
 OPTFLAGS = -march=native
 

Makefile.win

@@ -1,11 +1,10 @@
 EXTENSION = vector
-EXTVERSION = 0.5.1
+EXTVERSION = 0.4.4
 
 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)
+REGRESS_OPTS = --inputdir=test --load-extension=vector
 
 # For /arch flags
 # https://learn.microsoft.com/en-us/cpp/build/reference/arch-minimum-cpu-architecture
@@ -55,8 +54,6 @@ install:
 	copy $(SHLIB) "$(PKGLIBDIR)"
 	copy $(EXTENSION).control "$(SHAREDIR)\extension"
 	copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
-	mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
-	for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 installcheck:
 	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
@@ -64,9 +61,7 @@ installcheck:
 uninstall:
 	del /f "$(PKGLIBDIR)\$(SHLIB)"
 	del /f "$(SHAREDIR)\extension\$(EXTENSION).control"
-	del /f "$(SHAREDIR)\extension\$(EXTENSION)--*.sql"
+	del /f "$(SHAREDIR)\extension\vector--*.sql"
-	del /f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)\*.h"
-	rmdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 clean:
 	del /f $(SHLIB) $(EXTENSION).lib $(EXTENSION).exp

README.md

@@ -2,7 +2,7 @@
 
 Open-source vector similarity search for Postgres
 
-Store your vectors with the rest of your data. Supports:
+Store all of your application data in one place. Supports:
 
 - exact and approximate nearest neighbor search
 - L2 distance, inner product, and cosine distance
@@ -10,7 +10,7 @@ Store your vectors with the rest of your data. Supports:
 
 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/actions/workflows/build.yml/badge.svg)](https://github.com/pgvector/pgvector/actions)
+[![Build Status](https://github.com/pgvector/pgvector/workflows/build/badge.svg?branch=master)](https://github.com/pgvector/pgvector/actions)
 
 ## Installation
 
@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
 
 ```sh
 cd /tmp
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.4.4 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -26,7 +26,7 @@ make install # may need sudo
 
 See the [installation notes](#installation-notes) if you run into issues
 
-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). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
+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
 
@@ -157,105 +157,7 @@ SELECT category_id, AVG(embedding) FROM items GROUP BY category_id;
 
 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.
+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.
-
-Supported index types are:
-
-- [HNSW](#hnsw) - added in 0.5.0
-- [IVFFlat](#ivfflat)
-
-## HNSW
-
-An HNSW index creates a multilayer graph. It has better query performance than IVFFlat (in terms of speed-recall tradeoff), but has slower build times and uses more memory. Also, an index can be created without any data in the table since there isn’t a training step like IVFFlat.
-
-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);
-```
-
-A higher value of `ef_construction` provides better recall at the cost of index build time / insert speed.
-
-### 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;
-```
-
-### Index Build Time
-
-Indexes build significantly faster when the graph fits into `maintenance_work_mem`
-
-```sql
-SET maintenance_work_mem = '8GB';
-```
-
-A notice is shown when the graph no longer fits
-
-```text
-NOTICE:  hnsw graph no longer fits into maintenance_work_mem after 100000 tuples
-DETAIL:  Building will take significantly more time.
-HINT:  Increase maintenance_work_mem to speed up builds.
-```
-
-Note: Do not set `maintenance_work_mem` so high that it exhausts the memory on the server
-
-### Indexing Progress
-
-Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
-
-```sql
-SELECT phase, round(100.0 * blocks_done / nullif(blocks_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
-```
-
-The phases for HNSW are:
-
-1. `initializing`
-2. `loading tuples`
-
-## 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:
 
@@ -304,32 +206,22 @@ SELECT ...
 COMMIT;
 ```
 
-### Index Build Time
-
-Speed up index creation on large tables by increasing the number of parallel workers (2 by default)
-
-```sql
-SET max_parallel_maintenance_workers = 7; -- plus leader
-```
-
-For a large number of workers, you may also need to increase `max_parallel_workers` (8 by default)
-
 ### Indexing Progress
 
 Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
 
 ```sql
-SELECT phase, round(100.0 * tuples_done / nullif(tuples_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
+SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
 ```
 
-The phases for IVFFlat are:
+The phases are:
 
 1. `initializing`
 2. `performing k-means`
-3. `assigning tuples`
+3. `sorting tuples`
 4. `loading tuples`
 
-Note: `%` is only populated during the `loading tuples` phase
+Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
 
 ## Filtering
 
@@ -348,7 +240,8 @@ 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 hnsw (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);
 ```
 
 Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html) for approximate search on many different values of the `WHERE` columns
@@ -359,15 +252,13 @@ CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(cate
 
 ## Hybrid Search
 
-Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for 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;
 ```
 
-You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
-
 ## Performance
 
 Use `EXPLAIN ANALYZE` to debug performance.
@@ -392,7 +283,7 @@ 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).
+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);
@@ -404,21 +295,17 @@ Use pgvector from any language with a Postgres client. You can even generate and
 
 Language | Libraries / Examples
 --- | ---
-C | [pgvector-c](https://github.com/pgvector/pgvector-c)
 C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
-C#, F#, Visual Basic | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
+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, Kotlin, Groovy, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
+Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
-JavaScript, TypeScript | [pgvector-node](https://github.com/pgvector/pgvector-node)
 Julia | [pgvector-julia](https://github.com/pgvector/pgvector-julia)
-Lisp | [pgvector-lisp](https://github.com/pgvector/pgvector-lisp)
 Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
-Nim | [pgvector-nim](https://github.com/pgvector/pgvector-nim)
+Node.js | [pgvector-node](https://github.com/pgvector/pgvector-node)
-OCaml | [pgvector-ocaml](https://github.com/pgvector/pgvector-ocaml)
 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)
@@ -426,7 +313,6 @@ 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)
-Zig | [pgvector-zig](https://github.com/pgvector/pgvector-zig)
 
 ## Frequently Asked Questions
 
@@ -442,63 +328,6 @@ Yes, pgvector uses the write-ahead log (WAL), which allows for replication and p
 
 You’ll need to use [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction) at the moment.
 
-#### Can I store vectors with different dimensions in the same column?
-
-You can use `vector` as the type (instead of `vector(3)`).
-
-```sql
-CREATE TABLE embeddings (model_id bigint, item_id bigint, embedding vector, PRIMARY KEY (model_id, item_id));
-```
-
-However, you can only create indexes on rows with the same number of dimensions (using [expression](https://www.postgresql.org/docs/current/indexes-expressional.html) and [partial](https://www.postgresql.org/docs/current/indexes-partial.html) indexing):
-
-```sql
-CREATE INDEX ON embeddings USING hnsw ((embedding::vector(3)) vector_l2_ops) WHERE (model_id = 123);
-```
-
-and query with:
-
-```sql
-SELECT * FROM embeddings WHERE model_id = 123 ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
-```
-
-#### Can I store vectors with more precision?
-
-You can use the `double precision[]` or `numeric[]` type to store vectors with more precision.
-
-```sql
-CREATE TABLE items (id bigserial PRIMARY KEY, embedding double precision[]);
-
--- use {} instead of [] for Postgres arrays
-INSERT INTO items (embedding) VALUES ('{1,2,3}'), ('{4,5,6}');
-```
-
-Optionally, add a [check constraint](https://www.postgresql.org/docs/current/ddl-constraints.html) to ensure data can be converted to the `vector` type and has the expected dimensions.
-
-```sql
-ALTER TABLE items ADD CHECK (vector_dims(embedding::vector) = 3);
-```
-
-Use [expression indexing](https://www.postgresql.org/docs/current/indexes-expressional.html) to index (at a lower precision):
-
-```sql
-CREATE INDEX ON items USING hnsw ((embedding::vector(3)) vector_l2_ops);
-```
-
-and query with:
-
-```sql
-SELECT * FROM items ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
-```
-
-#### Do indexes need to fit into memory?
-
-No, but like other index types, you’ll likely see better performance if they do. You can get the size of an index with:
-
-```sql
-SELECT pg_size_pretty(pg_relation_size('index_name'));
-```
-
 ## Troubleshooting
 
 #### Why isn’t a query using an index?
@@ -512,8 +341,6 @@ SELECT ...
 COMMIT;
 ```
 
-Also, if the table is small, a table scan may be faster.
-
 #### 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:
@@ -532,7 +359,7 @@ or choose to store vectors inline:
 ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
 ```
 
-#### Why are there less results for a query after adding an IVFFlat index?
+#### Why are there less results for a query 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.
 
@@ -548,32 +375,33 @@ Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a singl
 
 ### Vector Operators
 
-Operator | Description | Added
+Operator | Description
---- | --- | ---
+--- | ---
-\+ | element-wise addition |
+\+ | element-wise addition
-\- | element-wise subtraction |
+\- | element-wise subtraction
-\* | element-wise multiplication | 0.5.0
+\* | element-wise multiplication [unreleased]
-<-> | Euclidean distance |
+<-> | Euclidean distance
-<#> | negative inner product |
+<#> | negative inner product
-<=> | cosine distance |
+<=> | cosine distance
 
 ### Vector Functions
 
-Function | Description | Added
+Function | Description
---- | --- | ---
+--- | ---
-cosine_distance(vector, vector) → double precision | cosine distance |
+cosine_distance(vector, vector) → double precision | cosine distance
-inner_product(vector, vector) → double precision | inner product |
+inner_product(vector, vector) → double precision | inner product
-l2_distance(vector, vector) → double precision | Euclidean distance |
+l2_distance(vector, vector) → double precision | Euclidean distance
-l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
+l1_distance(vector, vector) → double precision | taxicab distance [unreleased]
-vector_dims(vector) → integer | number of dimensions |
+normalize_l2(vector) → vector | normalize with Euclidean norm [unreleased]
-vector_norm(vector) → double precision | Euclidean norm |
+vector_dims(vector) → integer | number of dimensions
+vector_norm(vector) → double precision | Euclidean norm
 
 ### Aggregate Functions
 
-Function | Description | Added
+Function | Description
---- | --- | ---
+--- | ---
-avg(vector) → vector | average |
+avg(vector) → vector | arithmetic mean
-sum(vector) → vector | sum | 0.5.0
+sum(vector) → vector | sum [unreleased]
 
 ## Installation Notes
 
@@ -582,7 +410,7 @@ sum(vector) → vector | sum | 0.5.0
 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=/Library/PostgreSQL/16/bin/pg_config
+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:
@@ -591,14 +419,6 @@ Then re-run the installation instructions (run `make clean` before `make` if nee
 sudo --preserve-env=PG_CONFIG make install
 ```
 
-A few common paths on Mac are:
-
-- EDB installer - `/Library/PostgreSQL/16/bin/pg_config`
-- Homebrew (arm64) - `/opt/homebrew/opt/postgresql@16/bin/pg_config`
-- Homebrew (x86-64) - `/usr/local/opt/postgresql@16/bin/pg_config`
-
-Note: Replace `16` with your Postgres server version
-
 ### 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.
@@ -606,14 +426,10 @@ If compilation fails with `fatal error: postgres.h: No such file or directory`,
 For Ubuntu and Debian, use:
 
 ```sh
-sudo apt install postgresql-server-dev-16
+sudo apt install postgresql-server-dev-15
 ```
 
-Note: Replace `16` with your Postgres server version
+Note: Replace `15` with your Postgres server version
-
-### Missing SDK
-
-If compilation fails and the output includes `warning: no such sysroot directory` on Mac, reinstall Xcode Command Line Tools.
 
 ### Windows
 
@@ -628,8 +444,8 @@ Note: The exact path will vary depending on your Visual Studio version and editi
 Then use `nmake` to build:
 
 ```cmd
-set "PGROOT=C:\Program Files\PostgreSQL\16"
+set "PGROOT=C:\Program Files\PostgreSQL\15"
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.4.4 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -650,8 +466,9 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
+git clone --branch v0.4.4 https://github.com/pgvector/pgvector.git
 cd pgvector
+git cherry-pick 237a6df
 docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
 ```
 
@@ -678,22 +495,22 @@ pgxn install vector
 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-16-pgvector
+sudo apt install postgresql-15-pgvector
 ```
 
-Note: Replace `16` with your Postgres server version
+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_16
+sudo yum install pgvector_15
 # or
-sudo dnf install pgvector_16
+sudo dnf install pgvector_15
 ```
 
-Note: Replace `16` with your Postgres server version
+Note: Replace `15` with your Postgres server version
 
 ### conda-forge
 
@@ -715,18 +532,12 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
 
 ## Upgrading
 
-[Install](#installation) the latest version (use the same method as the original installation). Then in each database you want to upgrade, run:
+Install the latest version and run:
 
 ```sql
 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
@@ -803,25 +614,7 @@ make prove_installcheck PROVE_TESTS=test/t/001_wal.pl  # TAP test
 To enable benchmarking:
 
 ```sh
-make clean && PG_CFLAGS="-DIVFFLAT_BENCH" make && make install
+make clean && PG_CFLAGS=-DIVFFLAT_BENCH make && make install
-```
-
-To show memory usage:
-
-```sh
-make clean && PG_CFLAGS="-DHNSW_MEMORY -DIVFFLAT_MEMORY" make && make install
-```
-
-To enable assertions:
-
-```sh
-make clean && PG_CFLAGS="-DUSE_ASSERT_CHECKING" make && make install
-```
-
-To get k-means metrics:
-
-```sh
-make clean && PG_CFLAGS="-DIVFFLAT_KMEANS_DEBUG" make && make install
 ```
 
 Resources for contributors

sql/vector--0.4.4--0.5.0.sql

@@ -4,6 +4,9 @@
 CREATE FUNCTION l1_distance(vector, vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION normalize_l2(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;
 

sql/vector--0.5.0--0.5.1.sql

@@ -1,2 +0,0 @@
--- complain if script is sourced in psql, rather than via CREATE EXTENSION
-\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit

sql/vector.sql

@@ -49,6 +49,9 @@ CREATE FUNCTION vector_dims(vector) RETURNS integer
 CREATE FUNCTION vector_norm(vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION normalize_l2(vector) RETURNS vector
+	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;
 
@@ -265,15 +268,15 @@ CREATE OPERATOR CLASS vector_ip_ops
 	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);
+	FUNCTION 6 normalize_l2(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);
+	FUNCTION 5 normalize_l2(vector),
+	FUNCTION 6 normalize_l2(vector);
 
 CREATE OPERATOR CLASS vector_l2_ops
 	FOR TYPE vector USING hnsw AS
@@ -289,4 +292,4 @@ 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);
+	FUNCTION 3 normalize_l2(vector);

src/hnsw.c

@@ -14,7 +14,6 @@
 #endif
 
 int			hnsw_ef_search;
-bool		hnsw_enable_parallel_build;
 static relopt_kind hnsw_relopt_kind;
 
 /*
@@ -38,13 +37,8 @@ HnswInit(void)
 		);
 
 	DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
-							"Valid range is 1..1000.", &hnsw_ef_search,
+							"Valid range is 10..1000.", &hnsw_ef_search,
 							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
-
-	/* Behind a variable for now since can be slower than building in memory */
-	DefineCustomBoolVariable("hnsw.enable_parallel_build", "Enables or disables building indexes in parallel",
-							 NULL, &hnsw_enable_parallel_build,
-							 false, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
 /*
@@ -94,24 +88,10 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 		return;
 	}
 
-	/*
-	 * Do not use index if no limit or limit + offset > ef_search unless
-	 * enable_seqscan = off
-	 */
-	if (root->limit_tuples < 0 || root->limit_tuples > hnsw_ef_search)
-	{
-		*indexStartupCost = 1.0e10 - 1;
-		*indexTotalCost = 1.0e10 - 1;
-		*indexSelectivity = 0;
-		*indexCorrelation = 0;
-		*indexPages = 0;
-		return;
-	}
-
 	MemSet(&costs, 0, sizeof(costs));
 
 	index = index_open(path->indexinfo->indexoid, NoLock);
-	HnswGetMetaPageInfo(index, &m, NULL);
+	m = HnswGetM(index);
 	index_close(index, NoLock);
 
 	/* Approximate entry level */
@@ -187,7 +167,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 2;
+	amroutine->amsupport = 3;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif
@@ -216,7 +196,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 	amroutine->aminsert = hnswinsert;
 	amroutine->ambulkdelete = hnswbulkdelete;
 	amroutine->amvacuumcleanup = hnswvacuumcleanup;
-	amroutine->amcanreturn = NULL;
+	amroutine->amcanreturn = NULL;	/* tuple not included in heapsort */
 	amroutine->amcostestimate = hnswcostestimate;
 	amroutine->amoptions = hnswoptions;
 	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */

src/hnsw.h

@@ -4,9 +4,7 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
-#include "access/parallel.h"
 #include "access/reloptions.h"
-#include "lib/ilist.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
 #include "utils/sampling.h"
@@ -16,56 +14,42 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
-#if PG_VERSION_NUM < 120000
-#include "access/relscan.h"
-#endif
-
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
 #define HNSW_DISTANCE_PROC 1
 #define HNSW_NORM_PROC 2
+#define HNSW_NORMALIZE_PROC 3
 
 #define HNSW_VERSION	1
 #define HNSW_MAGIC_NUMBER 0xA953A953
-#define HNSW_PAGE_ID	0xFF90
+#define HNSW_PAGE_ID	0xFF85
 
 /* 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_MIN_M	4
 #define HNSW_MAX_M		100
-#define HNSW_DEFAULT_EF_CONSTRUCTION	64
+#define HNSW_DEFAULT_EF_CONSTRUCTION	40
-#define HNSW_MIN_EF_CONSTRUCTION	4
+#define HNSW_MIN_EF_CONSTRUCTION	10
 #define HNSW_MAX_EF_CONSTRUCTION		1000
 #define HNSW_DEFAULT_EF_SEARCH	40
-#define HNSW_MIN_EF_SEARCH		1
+#define HNSW_MIN_EF_SEARCH		10
 #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_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
+#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
-
-#define HNSW_ELEMENT_TUPLE_SIZE(size)	MAXALIGN(offsetof(HnswElementTupleData, data) + (size))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -73,10 +57,8 @@
 
 #if PG_VERSION_NUM >= 150000
 #define RandomDouble() pg_prng_double(&pg_global_prng_state)
-#define SeedRandom(seed) pg_prng_seed(&pg_global_prng_state, seed)
 #else
 #define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
-#define SeedRandom(seed) srandom(seed)
 #endif
 
 #if PG_VERSION_NUM < 130000
@@ -87,35 +69,28 @@
 #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)
-#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 */
+/* Ensure fits in uint8 */
-#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / (m)) - 2, 255)
+#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / m) - 2, 255)
-
-#define HnswGetNeighbors(element, lc) (AssertMacro((element)->level >= (lc)), &(element)->neighbors[lc])
 
 /* Variables */
 extern int	hnsw_ef_search;
-extern bool hnsw_enable_parallel_build;
+
+typedef struct HnswNeighborArray HnswNeighborArray;
 
 typedef struct HnswElementData
 {
-	slist_node	next;
+	List	   *heaptids;
-	ItemPointerData heaptids[HNSW_HEAPTIDS];
-	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
-	uint32		hash;
+	HnswNeighborArray *neighbors;
-	struct HnswNeighborArray *neighbors;
 	BlockNumber blkno;
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Datum		value;
+	Vector	   *vec;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -124,16 +99,21 @@ typedef struct HnswCandidate
 {
 	HnswElement element;
 	float		distance;
-	bool		closer;
 }			HnswCandidate;
 
 typedef struct HnswNeighborArray
 {
 	int			length;
-	bool		closerSet;
 	HnswCandidate *items;
 }			HnswNeighborArray;
 
+typedef struct HnswUpdate
+{
+	HnswCandidate hc;
+	int			level;
+	int			index;
+}			HnswUpdate;
+
 typedef struct HnswPairingHeapNode
 {
 	pairingheap_node ph_node;
@@ -148,59 +128,6 @@ typedef struct HnswOptions
 	int			efConstruction; /* size of dynamic candidate list */
 }			HnswOptions;
 
-typedef struct HnswGraph
-{
-	slist_head	elements;
-	HnswElement entryPoint;
-	long		memoryUsed;
-	long		memoryTotal;
-	bool		flushed;
-	double		indtuples;
-}			HnswGraph;
-
-typedef struct HnswSpool
-{
-	Relation	heap;
-	Relation	index;
-}			HnswSpool;
-
-typedef struct HnswShared
-{
-	/* 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;
-	HnswGraph	graphData;
-
-#if PG_VERSION_NUM < 120000
-	ParallelHeapScanDescData heapdesc;	/* must come last */
-#endif
-}			HnswShared;
-
-#if PG_VERSION_NUM >= 120000
-#define ParallelTableScanFromHnswShared(shared) \
-	(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(HnswShared)))
-#endif
-
-typedef struct HnswLeader
-{
-	ParallelContext *pcxt;
-	int			nparticipanttuplesorts;
-	HnswShared *hnswshared;
-	Snapshot	snapshot;
-}			HnswLeader;
-
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -221,22 +148,20 @@ typedef struct HnswBuildState
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
-	HnswGraph	graphData;
+	List	   *elements;
-	HnswGraph  *graph;
+	HnswElement entryPoint;
 	double		ml;
 	int			maxLevel;
+	double		maxInMemoryElements;
+	bool		flushed;
 	Vector	   *normvec;
 
 	/* Memory */
-	MemoryContext graphCtx;
 	MemoryContext tmpCtx;
-
-	/* Parallel builds */
-	HnswLeader *hnswleader;
-	HnswShared *hnswshared;
 }			HnswBuildState;
 
 typedef struct HnswMetaPageData
@@ -272,7 +197,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		data;
+	Vector		vec;
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -290,12 +215,14 @@ typedef HnswNeighborTupleData * HnswNeighborTuple;
 typedef struct HnswScanOpaqueData
 {
 	bool		first;
+	Buffer		buf;
 	List	   *w;
 	MemoryContext tmpCtx;
 
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 }			HnswScanOpaqueData;
 
@@ -318,7 +245,7 @@ typedef struct HnswVacuumState
 	Oid			collation;
 
 	/* Variables */
-	struct tidhash_hash *deleted;
+	HTAB	   *deleted;
 	BufferAccessStrategy bas;
 	HnswNeighborTuple ntup;
 	HnswElementData highestPoint;
@@ -330,30 +257,26 @@ typedef struct HnswVacuumState
 /* Methods */
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
-FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
+FmgrInfo   *HnswOptionalProcInfo(Relation rel, uint16 procnum);
-bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+void		HnswNormValue(FmgrInfo *procinfo, FmgrInfo *normalizeprocinfo, 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);
+List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, BlockNumber *skipPage, OffsetNumber *skipOffno);
 HnswElement HnswGetEntryPoint(Relation index);
-void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
-HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
+void		HnswFreeElement(HnswElement element);
-void		HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
+HnswElement HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, List **updates, bool vacuuming);
-HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
+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, bool building);
+void		HnswUpdateMetaPage(Relation index, bool 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, bool building);
+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, bool building);
-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);
-PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);
@@ -371,31 +294,4 @@ void		hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys,
 bool		hnswgettuple(IndexScanDesc scan, ScanDirection dir);
 void		hnswendscan(IndexScanDesc scan);
 
-/* Hash tables */
-typedef struct TidHashEntry
-{
-	ItemPointerData tid;
-	char		status;
-}			TidHashEntry;
-
-#define SH_PREFIX tidhash
-#define SH_ELEMENT_TYPE TidHashEntry
-#define SH_KEY_TYPE ItemPointerData
-#define SH_SCOPE extern
-#define SH_DECLARE
-#include "lib/simplehash.h"
-
-typedef struct PointerHashEntry
-{
-	uintptr_t	ptr;
-	char		status;
-}			PointerHashEntry;
-
-#define SH_PREFIX pointerhash
-#define SH_ELEMENT_TYPE PointerHashEntry
-#define SH_KEY_TYPE uintptr_t
-#define SH_SCOPE extern
-#define SH_DECLARE
-#include "lib/simplehash.h"
-
 #endif

src/hnswinsert.c

@@ -5,7 +5,6 @@
 #include "hnsw.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
-#include "utils/datum.h"
 #include "utils/memutils.h"
 
 /*
@@ -35,7 +34,7 @@ GetInsertPage(Relation index)
  * 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)
+HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *firstFreePage)
 {
 	OffsetNumber offno;
 	OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
@@ -50,15 +49,14 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
 
 		if (etup->deleted)
 		{
-			BlockNumber elementPage = BufferGetBlockNumber(buf);
 			BlockNumber neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
 			OffsetNumber neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
 			ItemId		itemid;
 
-			if (!BlockNumberIsValid(*newInsertPage))
+			if (!BlockNumberIsValid(*firstFreePage))
-				*newInsertPage = elementPage;
+				*firstFreePage = neighborPage;
 
-			if (neighborPage == elementPage)
+			if (neighborPage == BufferGetBlockNumber(buf))
 			{
 				*nbuf = buf;
 				*npage = page;
@@ -93,7 +91,7 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
  * Add a new page
  */
 static void
-HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page, bool building)
+HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page)
 {
 	/* Add a new page */
 	LockRelationForExtension(index, ExclusiveLock);
@@ -101,11 +99,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
 	UnlockRelationForExtension(index, ExclusiveLock);
 
 	/* Init new page */
-	if (building)
+	*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
-		*npage = BufferGetPage(*nbuf);
-	else
-		*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
-
 	HnswInitPage(*nbuf, *npage);
 
 	/* Update previous buffer */
@@ -116,7 +110,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
  * Add to element and neighbor pages
  */
 static void
-WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage, bool building)
+WriteNewElementPages(Relation index, HnswElement e, int m)
 {
 	Buffer		buf;
 	Page		page;
@@ -124,23 +118,21 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	Size		etupSize;
 	Size		ntupSize;
 	Size		combinedSize;
-	Size		maxSize;
-	Size		minCombinedSize;
 	HnswElementTuple etup;
-	BlockNumber currentPage = insertPage;
+	BlockNumber insertPage = GetInsertPage(index);
+	BlockNumber originalInsertPage = insertPage;
+	int			dimensions = e->vec->dim;
 	HnswNeighborTuple ntup;
 	Buffer		nbuf;
 	Page		npage;
 	OffsetNumber freeOffno = InvalidOffsetNumber;
 	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
-	BlockNumber newInsertPage = InvalidBlockNumber;
+	BlockNumber firstFreePage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
-	maxSize = HNSW_MAX_SIZE;
-	minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
 
 	/* Prepare element tuple */
 	etup = palloc0(etupSize);
@@ -150,30 +142,16 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	ntup = palloc0(ntupSize);
 	HnswSetNeighborTuple(ntup, e, m);
 
-	/* Find a page (or two if needed) to insert the tuples */
+	/* Find a page to insert the item */
 	for (;;)
 	{
-		buf = ReadBuffer(index, currentPage);
+		buf = ReadBuffer(index, insertPage);
 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
 
-		if (building)
+		state = GenericXLogStart(index);
-		{
+		page = GenericXLogRegisterBuffer(state, buf, 0);
-			state = NULL;
-			page = BufferGetPage(buf);
-		}
-		else
-		{
-			state = GenericXLogStart(index);
-			page = GenericXLogRegisterBuffer(state, buf, 0);
-		}
 
-		/* Keep track of first page where element at level 0 can fit */
+		/* Space for both */
-		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;
@@ -181,35 +159,28 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 			break;
 		}
 
-		/* Next, try space from a deleted element */
+		/* Space for element but not neighbors and last page */
-		if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
+		if (PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
+		{
+			HnswInsertAppendPage(index, &nbuf, &npage, state, page);
+			break;
+		}
+
+		/* Space from deleted item */
+		if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &firstFreePage))
 		{
 			if (nbuf != buf)
-			{
+				npage = GenericXLogRegisterBuffer(state, nbuf, 0);
-				if (building)
-					npage = BufferGetPage(nbuf);
-				else
-					npage = GenericXLogRegisterBuffer(state, nbuf, 0);
-			}
 
 			break;
 		}
 
-		/* Finally, try space for element only if last page */
+		insertPage = HnswPageGetOpaque(page)->nextblkno;
-		/* 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, building);
-			break;
-		}
 
-		currentPage = HnswPageGetOpaque(page)->nextblkno;
+		if (BlockNumberIsValid(insertPage))
-
-		if (BlockNumberIsValid(currentPage))
 		{
 			/* Move to next page */
-			if (!building)
+			GenericXLogAbort(state);
-				GenericXLogAbort(state);
 			UnlockReleaseBuffer(buf);
 		}
 		else
@@ -217,33 +188,24 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 			Buffer		newbuf;
 			Page		newpage;
 
-			HnswInsertAppendPage(index, &newbuf, &newpage, state, page, building);
+			HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
 
 			/* Commit */
-			if (building)
+			MarkBufferDirty(newbuf);
-				MarkBufferDirty(buf);
+			MarkBufferDirty(buf);
-			else
+			GenericXLogFinish(state);
-				GenericXLogFinish(state);
 
 			/* Unlock previous buffer */
 			UnlockReleaseBuffer(buf);
 
 			/* Prepare new buffer */
+			state = GenericXLogStart(index);
 			buf = newbuf;
-			if (building)
+			page = GenericXLogRegisterBuffer(state, buf, 0);
-			{
-				state = NULL;
-				page = BufferGetPage(buf);
-			}
-			else
-			{
-				state = GenericXLogStart(index);
-				page = GenericXLogRegisterBuffer(state, buf, 0);
-			}
 
 			/* Create new page for neighbors if needed */
 			if (PageGetFreeSpace(page) < combinedSize)
-				HnswInsertAppendPage(index, &nbuf, &npage, state, page, building);
+				HnswInsertAppendPage(index, &nbuf, &npage, state, page);
 			else
 			{
 				nbuf = buf;
@@ -257,10 +219,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	e->blkno = BufferGetBlockNumber(buf);
 	e->neighborPage = BufferGetBlockNumber(nbuf);
 
-	/* Added tuple to new page if newInsertPage is not set */
+	insertPage = e->neighborPage;
-	/* So can set to neighbor page instead of element page */
-	if (!BlockNumberIsValid(newInsertPage))
-		newInsertPage = e->neighborPage;
 
 	if (OffsetNumberIsValid(freeOffno))
 	{
@@ -297,149 +256,84 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	}
 
 	/* Commit */
-	if (building)
+	MarkBufferDirty(buf);
-	{
+	if (nbuf != buf)
-		MarkBufferDirty(buf);
+		MarkBufferDirty(nbuf);
-		if (nbuf != buf)
+	GenericXLogFinish(state);
-			MarkBufferDirty(nbuf);
-	}
-	else
-		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 	if (nbuf != buf)
 		UnlockReleaseBuffer(nbuf);
 
 	/* Update the insert page */
-	if (BlockNumberIsValid(newInsertPage) && newInsertPage != insertPage)
+	if (insertPage != originalInsertPage && (!OffsetNumberIsValid(freeOffno) || firstFreePage == insertPage))
-		*updatedInsertPage = newInsertPage;
+		HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
 }
 
 /*
- * Check if connection already exists
+ * Calculate index for update
  */
-static bool
+static int
-ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
+HnswGetIndex(HnswUpdate * update, int m)
 {
-	for (int i = 0; i < lm; i++)
+	return (update->hc.element->level - update->level) * m + update->index;
-	{
-		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
+static void
-HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
+UpdateNeighborPages(Relation index, HnswElement e, int m, List *updates)
 {
-	for (int lc = e->level; lc >= 0; lc--)
+	ListCell   *lc;
+
+	/* Could update multiple at once for same element */
+	/* but should only happen a low percent of time, so keep simple for now */
+	foreach(lc, updates)
 	{
-		int			lm = HnswGetLayerM(m, lc);
+		Buffer		buf;
-		HnswNeighborArray *neighbors = HnswGetNeighbors(e, lc);
+		Page		page;
+		GenericXLogState *state;
+		HnswUpdate *update = lfirst(lc);
+		ItemId		itemid;
+		HnswNeighborTuple ntup;
+		Size		ntupSize;
+		int			idx;
+		OffsetNumber offno = update->hc.element->neighborOffno;
 
-		for (int i = 0; i < neighbors->length; i++)
+		/* Register page */
+		buf = ReadBuffer(index, update->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 */
+		idx = HnswGetIndex(update, m);
+
+		/* Make robust to issues */
+		if (idx < ntup->count)
 		{
-			HnswCandidate *hc = &neighbors->items[i];
+			ItemPointer indextid = &ntup->indextids[idx];
-			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 */
+			/* Update neighbor */
-			/* Do not lock yet since selecting neighbors can take time */
+			ItemPointerSet(indextid, e->blkno, e->offno);
-			HnswLoadNeighbors(hc->element, index, m);
 
-			/*
+			/* Overwrite tuple */
-			 * Could improve performance for vacuuming by checking neighbors
+			if (!PageIndexTupleOverwrite(page, offno, (Item) ntup, ntupSize))
-			 * against list of elements being deleted to find index. It's
+				elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
-			 * important to exclude already deleted elements for this since
-			 * they can be replaced at any time.
-			 */
 
-			/* Select neighbors */
+			/* Commit */
-			HnswUpdateConnection(e, hc, lm, lc, &idx, index, procinfo, collation);
+			MarkBufferDirty(buf);
-
+			GenericXLogFinish(state);
-			/* 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);
-			if (building)
-			{
-				state = NULL;
-				page = BufferGetPage(buf);
-			}
-			else
-			{
-				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 */
-				if (building)
-					MarkBufferDirty(buf);
-				else
-					GenericXLogFinish(state);
-			}
-			else if (!building)
-				GenericXLogAbort(state);
-
-			UnlockReleaseBuffer(buf);
 		}
+		else
+			GenericXLogAbort(state);
+
+		UnlockReleaseBuffer(buf);
 	}
 }
 
@@ -447,34 +341,23 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
  * Add a heap TID to an existing element
  */
 static bool
-HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool building)
+HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 {
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	ItemId		itemid;
+	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
 	HnswElementTuple etup;
-	Size		etupSize;
 	int			i;
 
 	/* Read page */
 	buf = ReadBuffer(index, dup->blkno);
 	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-	if (building)
+	state = GenericXLogStart(index);
-	{
+	page = GenericXLogRegisterBuffer(state, buf, 0);
-		state = NULL;
-		page = BufferGetPage(buf);
-	}
-	else
-	{
-		state = GenericXLogStart(index);
-		page = GenericXLogRegisterBuffer(state, buf, 0);
-	}
 
 	/* Find space */
-	itemid = PageGetItemId(page, dup->offno);
+	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
-	etup = (HnswElementTuple) PageGetItem(page, itemid);
-	etupSize = ItemIdGetLength(itemid);
 	for (i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (!ItemPointerIsValid(&etup->heaptids[i]))
@@ -484,142 +367,87 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool buil
 	/* Either being deleted or we lost our chance to another backend */
 	if (i == 0 || i == HNSW_HEAPTIDS)
 	{
-		if (!building)
+		GenericXLogAbort(state);
-			GenericXLogAbort(state);
 		UnlockReleaseBuffer(buf);
 		return false;
 	}
 
 	/* Add heap TID */
-	etup->heaptids[i] = element->heaptids[0];
+	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 */
-	if (building)
+	MarkBufferDirty(buf);
-		MarkBufferDirty(buf);
+	GenericXLogFinish(state);
-	else
-		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 
 	return true;
 }
 
-/*
- * Find duplicate element
- */
-static bool
-HnswFindDuplicate(Relation index, HnswElement element, bool building)
-{
-	HnswNeighborArray *neighbors = HnswGetNeighbors(element, 0);
-
-	for (int i = 0; i < neighbors->length; i++)
-	{
-		HnswCandidate *neighbor = &neighbors->items[i];
-
-		/* Exit early since ordered by distance */
-		if (!datumIsEqual(element->value, neighbor->element->value, false, -1))
-			return false;
-
-		if (HnswAddDuplicate(index, element, neighbor->element, building))
-			return true;
-	}
-
-	return false;
-}
-
 /*
  * Write changes to disk
  */
 static void
-WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
+WriteElement(Relation index, HnswElement element, int m, List *updates, HnswElement dup, HnswElement entryPoint)
 {
-	BlockNumber newInsertPage = InvalidBlockNumber;
+	/* Try to add to existing page */
+	if (dup != NULL)
+	{
+		if (HnswAddDuplicate(index, element, dup))
+			return;
+	}
 
-	/* Look for duplicate */
+	/* If fails, take this path */
-	if (HnswFindDuplicate(index, element, building))
+	WriteNewElementPages(index, element, m);
-		return;
+	UpdateNeighborPages(index, element, m, updates);
 
-	/* Write element and neighbor tuples */
+	/* Update metapage if needed */
-	WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage, building);
-
-	/* Update insert page if needed */
-	if (BlockNumberIsValid(newInsertPage))
-		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);
-
-	/* Update neighbors */
-	HnswUpdateNeighborPages(index, procinfo, collation, element, m, false, building);
-
-	/* Update entry point if needed */
 	if (entryPoint == NULL || element->level > entryPoint->level)
-		HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, building);
+		HnswUpdateMetaPage(index, true, element, InvalidBlockNumber, MAIN_FORKNUM);
 }
 
 /*
  * Insert a tuple into the index
  */
 bool
-HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building)
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
 {
 	Datum		value;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	HnswElement entryPoint;
 	HnswElement element;
-	int			m;
+	int			m = HnswGetM(index);
 	int			efConstruction = HnswGetEfConstruction(index);
+	double		ml = HnswGetMl(m);
 	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	Oid			collation = index->rd_indcollation[0];
-	LOCKMODE	lockmode = ShareLock;
+	List	   *updates = NIL;
+	HnswElement dup;
 
 	/* Detoast once for all calls */
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
-	if (normprocinfo != NULL)
+	normalizeprocinfo = HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC);
-	{
+	HnswNormValue(normprocinfo, normalizeprocinfo, collation, &value, 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 = HnswInitElement(heap_tid, m, ml, HnswGetMaxLevel(m));
-	element->value = value;
+	element->vec = DatumGetVector(value);
 
-	/* Prevent concurrent inserts when likely updating entry point */
+	/* Get entry point */
-	if (entryPoint == NULL || element->level > entryPoint->level)
+	entryPoint = HnswGetEntryPoint(index);
-	{
-		/* 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);
+	dup = HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, &updates, false);
 
 	/* Write to disk */
-	WriteElement(index, procinfo, collation, element, m, efConstruction, entryPoint, building);
+	WriteElement(index, element, m, updates, dup, entryPoint);
-
-	/* Release lock */
-	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
 
 	return true;
 }
@@ -650,7 +478,7 @@ hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 	oldCtx = MemoryContextSwitchTo(insertCtx);
 
 	/* Insert tuple */
-	HnswInsertTuple(index, values, isnull, heap_tid, heap, false);
+	HnswInsertTuple(index, values, isnull, heap_tid, heap);
 
 	/* Delete memory context */
 	MemoryContextSwitchTo(oldCtx);

src/hnswscan.c

@@ -4,39 +4,34 @@
 #include "hnsw.h"
 #include "pgstat.h"
 #include "storage/bufmgr.h"
-#include "storage/lmgr.h"
 #include "utils/memutils.h"
 
 /*
  * Algorithm 5 from paper
  */
-static List *
+static void
 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	   *ep = NIL;
 	List	   *w;
-	int			m;
+	HnswElement entryPoint = HnswGetEntryPoint(index);
-	HnswElement entryPoint;
-
-	/* Get m and entry point */
-	HnswGetMetaPageInfo(index, &m, &entryPoint);
 
 	if (entryPoint == NULL)
-		return NIL;
+		return;
 
-	ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
+	ep = lappend(ep, 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);
+		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, false, NULL, NULL);
 		ep = w;
 	}
 
-	return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
+	so->w = HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, false, NULL, NULL);
 }
 
 /*
@@ -62,33 +57,6 @@ GetDimensions(Relation index)
 	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
  */
@@ -101,6 +69,7 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
 
 	so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
+	so->buf = InvalidBuffer;
 	so->first = true;
 	so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 									   "Hnsw scan temporary context",
@@ -109,6 +78,7 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	so->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	so->normalizeprocinfo = HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC);
 	so->collation = index->rd_indcollation[0];
 
 	scan->opaque = so;
@@ -160,50 +130,61 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan hnsw index without order");
 
-		/* Requires MVCC-compliant snapshot as not able to maintain a pin */
+		if (scan->orderByData->sk_flags & SK_ISNULL)
-		/* https://www.postgresql.org/docs/current/index-locking.html */
+			value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
-		if (!IsMVCCSnapshot(scan->xs_snapshot))
+		else
-			elog(ERROR, "non-MVCC snapshots are not supported with hnsw");
+		{
+			value = scan->orderByData->sk_argument;
 
-		/* Get scan value */
+			/* Value should not be compressed or toasted */
-		value = GetScanValue(scan);
+			Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+			Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
 
-		/*
+			/* Fine if normalization fails */
-		 * Get a shared lock. This allows vacuum to ensure no in-flight scans
+			HnswNormValue(so->normprocinfo, so->normalizeprocinfo, so->collation, &value, NULL);
-		 * before marking tuples as deleted.
+		}
-		 */
-		LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
-
-		so->w = GetScanItems(scan, value);
-
-		/* Release shared lock */
-		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
 
+		GetScanItems(scan, value);
 		so->first = false;
 	}
 
 	while (list_length(so->w) > 0)
 	{
 		HnswCandidate *hc = llast(so->w);
-		ItemPointer heaptid;
+		ItemPointer tid;
+		BlockNumber indexblkno;
 
-		/* Move to next element if no valid heap TIDs */
+		/* Move to next element if no valid heap tids */
-		if (hc->element->heaptidsLength == 0)
+		if (list_length(hc->element->heaptids) == 0)
 		{
 			so->w = list_delete_last(so->w);
 			continue;
 		}
 
-		heaptid = &hc->element->heaptids[--hc->element->heaptidsLength];
+		tid = llast(hc->element->heaptids);
+		indexblkno = hc->element->blkno;
+
+		hc->element->heaptids = list_delete_last(hc->element->heaptids);
 
 		MemoryContextSwitchTo(oldCtx);
 
 #if PG_VERSION_NUM >= 120000
-		scan->xs_heaptid = *heaptid;
+		scan->xs_heaptid = *tid;
 #else
-		scan->xs_ctup.t_self = *heaptid;
+		scan->xs_ctup.t_self = *tid;
 #endif
 
+		if (BufferIsValid(so->buf))
+			ReleaseBuffer(so->buf);
+
+		/*
+		 * An index scan must maintain a pin on the index page holding the
+		 * item last returned by amgettuple
+		 *
+		 * https://www.postgresql.org/docs/current/index-locking.html
+		 */
+		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
+
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -220,6 +201,10 @@ hnswendscan(IndexScanDesc scan)
 {
 	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
 
+	/* Release pin */
+	if (BufferIsValid(so->buf))
+		ReleaseBuffer(so->buf);
+
 	MemoryContextDelete(so->tmpCtx);
 
 	pfree(so);

src/hnswvacuum.c

@@ -5,16 +5,18 @@
 #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
+ * Check if deleted list contains an index tid
  */
 static bool
-DeletedContains(tidhash_hash * deleted, ItemPointer indextid)
+DeletedContains(HTAB *deleted, ItemPointer indextid)
 {
-	return tidhash_lookup(deleted, *indextid) != NULL;
+	bool		found;
+
+	hash_search(deleted, indextid, HASH_FIND, &found);
+	return found;
 }
 
 /*
@@ -30,7 +32,6 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 	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;
@@ -59,8 +60,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		/* Iterate over nodes */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			int			idx = 0;
 			bool		itemUpdated = false;
 
@@ -77,21 +77,17 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 						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 = ItemIdGetLength(itemid);
+					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -107,15 +103,13 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 			if (!ItemPointerIsValid(&etup->heaptids[0]))
 			{
 				ItemPointerData ip;
-				bool		found;
 
 				/* Add to deleted list */
 				ItemPointerSet(&ip, blkno, offno);
 
-				tidhash_insert(vacuumstate->deleted, ip, &found);
+				(void) hash_search(vacuumstate->deleted, &ip, HASH_ENTER, NULL);
-				Assert(!found);
 			}
-			else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
+			else if (etup->level > highestLevel && !(blkno == entryPoint->blkno && offno == entryPoint->offno))
 			{
 				/* Keep track of highest non-entry point */
 				highestPoint->blkno = blkno;
@@ -128,7 +122,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		blkno = HnswPageGetOpaque(page)->nextblkno;
 
 		if (updated)
+		{
+			MarkBufferDirty(buf);
 			GenericXLogFinish(state);
+		}
 		else
 			GenericXLogAbort(state);
 
@@ -172,11 +169,6 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
 		}
 	}
 
-	/* 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;
@@ -186,7 +178,7 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
  * Repair graph for a single element
  */
 static void
-RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
+RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element)
 {
 	Relation	index = vacuumstate->index;
 	Buffer		buf;
@@ -196,20 +188,42 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	int			efConstruction = vacuumstate->efConstruction;
 	FmgrInfo   *procinfo = vacuumstate->procinfo;
 	Oid			collation = vacuumstate->collation;
+	HnswElement entryPoint;
 	BufferAccessStrategy bas = vacuumstate->bas;
 	HnswNeighborTuple ntup = vacuumstate->ntup;
 	Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
 
-	/* Skip if element is entry point */
+	/* Check if any neighbors point to deleted values */
-	if (entryPoint != NULL && element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
+	if (!NeedsUpdated(vacuumstate, element))
 		return;
 
+	/* Refresh entry point for each element */
+	entryPoint = HnswGetEntryPoint(index);
+
+	/* Special case for entry point */
+	if (element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
+	{
+		if (BlockNumberIsValid(vacuumstate->highestPoint.blkno))
+		{
+			/* Already updated */
+			if (vacuumstate->highestPoint.blkno == element->blkno && vacuumstate->highestPoint.offno == element->offno)
+				return;
+
+			entryPoint = &vacuumstate->highestPoint;
+
+			/* Reset neighbors from previous update */
+			entryPoint->neighbors = NULL;
+		}
+		else
+			entryPoint = NULL;
+	}
+
 	/* Init fields */
 	HnswInitNeighbors(element, m);
-	element->heaptidsLength = 0;
+	element->heaptids = NIL;
 
 	/* Add element to graph, skipping itself */
-	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, true);
+	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, NULL, true);
 
 	/* Update neighbor tuple */
 	/* Do this before getting page to minimize locking */
@@ -226,11 +240,9 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 		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, false);
 }
 
 /*
@@ -244,35 +256,16 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 	HnswElement entryPoint;
 	MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
 
-	if (!BlockNumberIsValid(highestPoint->blkno))
+	/* Repair graph for highest non-entry point */
-		highestPoint = NULL;
+	/* This may not be the highest with new inserts, but should be fine */
-
+	if (BlockNumberIsValid(highestPoint->blkno))
-	/*
-	 * 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);
-
+		RepairGraphElement(vacuumstate, highestPoint);
-		/* 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 */
+	/* See if entry point needs updated */
-	LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
-
-	/* Get latest entry point */
 	entryPoint = HnswGetEntryPoint(index);
-
 	if (entryPoint != NULL)
 	{
 		ItemPointerData epData;
@@ -280,37 +273,15 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		ItemPointerSet(&epData, entryPoint->blkno, entryPoint->offno);
 
 		if (DeletedContains(vacuumstate->deleted, &epData))
-		{
+			HnswUpdateMetaPage(index, true, highestPoint, InvalidBlockNumber, MAIN_FORKNUM);
-			/*
-			 * 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, false);
-		}
 		else
 		{
-			/*
+			/* Highest point will be used to repair */
-			 * 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);
-
+			RepairGraphElement(vacuumstate, entryPoint);
-			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);
@@ -326,14 +297,6 @@ RepairGraph(HnswVacuumState * vacuumstate)
 	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))
@@ -370,8 +333,14 @@ RepairGraph(HnswVacuumState * vacuumstate)
 				continue;
 
 			/* Create an element */
-			element = HnswInitElementFromBlock(blkno, offno);
+			element = palloc(sizeof(HnswElementData));
-			HnswLoadElementFromTuple(element, etup, false, true);
+			element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
+			element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
+			element->level = etup->level;
+			element->blkno = blkno;
+			element->offno = offno;
+			element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
+			memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
 
 			elements = lappend(elements, element);
 		}
@@ -382,48 +351,7 @@ RepairGraph(HnswVacuumState * vacuumstate)
 
 		/* Update neighbor pages */
 		foreach(lc2, elements)
-		{
+			RepairGraphElement(vacuumstate, (HnswElement) lfirst(lc2));
-			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, false);
-
-			/* Release lock */
-			UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
-		}
 
 		/* Reset memory context */
 		MemoryContextSwitchTo(oldCtx);
@@ -441,14 +369,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	BlockNumber insertPage = InvalidBlockNumber;
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
-
+	IndexBulkDeleteResult *stats = vacuumstate->stats;
-	/*
-	 * 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))
 	{
@@ -477,8 +398,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		/* Update element and neighbors together */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
@@ -493,20 +413,20 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* 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]))
+			{
+				stats->num_index_tuples++;
 				continue;
+			}
+
+			/* Update stats */
+			stats->tuples_removed++;
 
 			/* Calculate sizes */
-			etupSize = ItemIdGetLength(itemid);
+			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -529,7 +449,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
+			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)
@@ -544,6 +464,9 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				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);
@@ -563,8 +486,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		UnlockReleaseBuffer(buf);
 	}
 
-	/* Update insert page last, after everything has been marked as deleted */
+	HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
-	HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM, false);
 }
 
 /*
@@ -574,6 +496,7 @@ 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));
@@ -582,6 +505,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	vacuumstate->stats = stats;
 	vacuumstate->callback = callback;
 	vacuumstate->callback_state = callback_state;
+	vacuumstate->m = HnswGetM(index);
 	vacuumstate->efConstruction = HnswGetEfConstruction(index);
 	vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
 	vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
@@ -591,11 +515,11 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 												"Hnsw vacuum temporary context",
 												ALLOCSET_DEFAULT_SIZES);
 
-	/* Get m from metapage */
-	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
-
 	/* Create hash table */
-	vacuumstate->deleted = tidhash_create(CurrentMemoryContext, 256, NULL);
+	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);
 }
 
 /*
@@ -604,7 +528,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 static void
 FreeVacuumState(HnswVacuumState * vacuumstate)
 {
-	tidhash_destroy(vacuumstate->deleted);
+	hash_destroy(vacuumstate->deleted);
 	FreeAccessStrategy(vacuumstate->bas);
 	pfree(vacuumstate->ntup);
 	MemoryContextDelete(vacuumstate->tmpCtx);

src/ivfbuild.c

@@ -10,8 +10,8 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
-#include "tcop/tcopprot.h"
 #include "utils/memutils.h"
+#include "tcop/tcopprot.h"
 
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"
@@ -75,11 +75,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 * Normalize with KMEANS_NORM_PROC since spherical distance function
 	 * expects unit vectors
 	 */
-	if (buildstate->kmeansnormprocinfo != NULL)
+	IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->kmeansnormalizeprocinfo, buildstate->collation, &value, buildstate->normvec);
-	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->normvec))
-			return;
-	}
 
 	if (samples->length < targsamples)
 	{
@@ -176,11 +172,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
-	if (buildstate->normprocinfo != NULL)
+	IvfflatNormValue(buildstate->normprocinfo, buildstate->normalizeprocinfo, buildstate->collation, &value, buildstate->normvec);
-	{
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
-			return;
-	}
 
 	/* Find the list that minimizes the distance */
 	for (int i = 0; i < centers->length; i++)
@@ -368,6 +360,8 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	buildstate->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	buildstate->kmeansnormprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
+	buildstate->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
+	buildstate->kmeansnormalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORMALIZE_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
 	/* Require more than one dimension for spherical k-means */
@@ -506,30 +500,29 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		listSize;
+	OffsetNumber offno;
+	Size		itemsz;
 	IvfflatList list;
 
-	listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
+	itemsz = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
-	list = palloc(listSize);
+	list = palloc(itemsz);
 
 	buf = IvfflatNewBuffer(index, forkNum);
 	IvfflatInitRegisterPage(index, &buf, &page, &state);
 
 	for (int i = 0; i < lists; i++)
 	{
-		OffsetNumber offno;
-
 		/* Load list */
 		list->startPage = InvalidBlockNumber;
 		list->insertPage = InvalidBlockNumber;
 		memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
 
 		/* Ensure free space */
-		if (PageGetFreeSpace(page) < listSize)
+		if (PageGetFreeSpace(page) < itemsz)
 			IvfflatAppendPage(index, &buf, &page, &state, forkNum);
 
 		/* Add the item */
-		offno = PageAddItem(page, (Item) list, listSize, InvalidOffsetNumber, false, false);
+		offno = PageAddItem(page, (Item) list, itemsz, InvalidOffsetNumber, false, false);
 		if (offno == InvalidOffsetNumber)
 			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
@@ -543,10 +536,10 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	pfree(list);
 }
 
-#ifdef IVFFLAT_KMEANS_DEBUG
 /*
  * Print k-means metrics
  */
+#ifdef IVFFLAT_KMEANS_DEBUG
 static void
 PrintKmeansMetrics(IvfflatBuildState * buildstate)
 {

src/ivfflat.c

@@ -24,7 +24,7 @@ IvfflatInit(void)
 {
 	ivfflat_relopt_kind = add_reloption_kind();
 	add_int_reloption(ivfflat_relopt_kind, "lists", "Number of inverted lists",
-					  IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS
+					  IVFFLAT_DEFAULT_LISTS, 1, IVFFLAT_MAX_LISTS
 #if PG_VERSION_NUM >= 130000
 					  ,AccessExclusiveLock
 #endif
@@ -32,7 +32,7 @@ IvfflatInit(void)
 
 	DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
 							"Valid range is 1..lists.", &ivfflat_probes,
-							IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
+							1, 1, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
 /*
@@ -71,7 +71,7 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	int			lists;
 	double		ratio;
 	double		spc_seq_page_cost;
-	Relation	index;
+	Relation	indexRel;
 #if PG_VERSION_NUM < 120000
 	List	   *qinfos;
 #endif
@@ -89,9 +89,9 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 
 	MemSet(&costs, 0, sizeof(costs));
 
-	index = index_open(path->indexinfo->indexoid, NoLock);
+	indexRel = index_open(path->indexinfo->indexoid, NoLock);
-	IvfflatGetMetaPageInfo(index, &lists, NULL);
+	lists = IvfflatGetLists(indexRel);
-	index_close(index, NoLock);
+	index_close(indexRel, NoLock);
 
 	/* Get the ratio of lists that we need to visit */
 	ratio = ((double) ivfflat_probes) / lists;
@@ -105,20 +105,6 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	 */
 	costs.numIndexTuples = path->indexinfo->tuples * ratio;
 
-	/*
-	 * Do not use index if no limit or limit + offset > expected tuples unless
-	 * enable_seqscan = off
-	 */
-	if (root->limit_tuples < 0 || root->limit_tuples > costs.numIndexTuples)
-	{
-		*indexStartupCost = 1.0e10 - 1;
-		*indexTotalCost = 1.0e10 - 1;
-		*indexSelectivity = 0;
-		*indexCorrelation = 0;
-		*indexPages = 0;
-		return;
-	}
-
 #if PG_VERSION_NUM >= 120000
 	genericcostestimate(root, path, loop_count, &costs);
 #else
@@ -208,7 +194,7 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 4;
+	amroutine->amsupport = 6;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif

src/ivfflat.h

@@ -31,6 +31,8 @@
 #define IVFFLAT_NORM_PROC 2
 #define IVFFLAT_KMEANS_DISTANCE_PROC 3
 #define IVFFLAT_KMEANS_NORM_PROC 4
+#define IVFFLAT_NORMALIZE_PROC 5
+#define IVFFLAT_KMEANS_NORMALIZE_PROC 6
 
 #define IVFFLAT_VERSION	1
 #define IVFFLAT_MAGIC_NUMBER 0x14FF1A7
@@ -40,11 +42,8 @@
 #define IVFFLAT_METAPAGE_BLKNO	0
 #define IVFFLAT_HEAD_BLKNO		1	/* first list page */
 
-/* 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 */
@@ -175,6 +174,8 @@ typedef struct IvfflatBuildState
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
 	FmgrInfo   *kmeansnormprocinfo;
+	FmgrInfo   *normalizeprocinfo;
+	FmgrInfo   *kmeansnormalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -244,8 +245,8 @@ typedef struct IvfflatScanList
 typedef struct IvfflatScanOpaqueData
 {
 	int			probes;
-	int			dimensions;
 	bool		first;
+	Buffer		buf;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -256,6 +257,7 @@ typedef struct IvfflatScanOpaqueData
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Lists */
@@ -275,10 +277,9 @@ 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 index, uint16 procnum);
+FmgrInfo   *IvfflatOptionalProcInfo(Relation rel, uint16 procnum);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+void		IvfflatNormValue(FmgrInfo *procinfo, FmgrInfo *normalizeprocinfo, Oid collation, Datum *value, Vector * result);
 int			IvfflatGetLists(Relation index);
-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);

src/ivfinsert.c

@@ -11,37 +11,36 @@
  * Find the list that minimizes the distance function
  */
 static void
-FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
+FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
 {
+	Buffer		cbuf;
+	Page		cpage;
+	IvfflatList list;
+	double		distance;
 	double		minDistance = DBL_MAX;
 	BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
 	FmgrInfo   *procinfo;
 	Oid			collation;
+	OffsetNumber offno;
+	OffsetNumber maxoffno;
 
 	/* Avoid compiler warning */
 	listInfo->blkno = nextblkno;
 	listInfo->offno = FirstOffsetNumber;
 
-	procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
+	procinfo = index_getprocinfo(rel, 1, IVFFLAT_DISTANCE_PROC);
-	collation = index->rd_indcollation[0];
+	collation = rel->rd_indcollation[0];
 
 	/* Search all list pages */
 	while (BlockNumberIsValid(nextblkno))
 	{
-		Buffer		cbuf;
+		cbuf = ReadBuffer(rel, nextblkno);
-		Page		cpage;
-		OffsetNumber maxoffno;
-
-		cbuf = ReadBuffer(index, nextblkno);
 		LockBuffer(cbuf, BUFFER_LOCK_SHARE);
 		cpage = BufferGetPage(cbuf);
 		maxoffno = PageGetMaxOffsetNumber(cpage);
 
-		for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			IvfflatList list;
-			double		distance;
-
 			list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
 			distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, values[0], PointerGetDatum(&list->center)));
 
@@ -64,11 +63,12 @@ FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo
  * Insert a tuple into the index
  */
 static void
-InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
+InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
 {
 	IndexTuple	itup;
 	Datum		value;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
@@ -81,33 +81,30 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
-	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
+	normprocinfo = IvfflatOptionalProcInfo(rel, IVFFLAT_NORM_PROC);
-	if (normprocinfo != NULL)
+	normalizeprocinfo = IvfflatOptionalProcInfo(rel, IVFFLAT_NORMALIZE_PROC);
-	{
+	IvfflatNormValue(normprocinfo, normalizeprocinfo, rel->rd_indcollation[0], &value, NULL);
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
-			return;
-	}
 
 	/* Find the insert page - sets the page and list info */
-	FindInsertPage(index, values, &insertPage, &listInfo);
+	FindInsertPage(rel, values, &insertPage, &listInfo);
 	Assert(BlockNumberIsValid(insertPage));
 	originalInsertPage = insertPage;
 
 	/* Form tuple */
-	itup = index_form_tuple(RelationGetDescr(index), &value, isnull);
+	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)) - sizeof(ItemIdData));
+	Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
 
 	/* Find a page to insert the item */
 	for (;;)
 	{
-		buf = ReadBuffer(index, insertPage);
+		buf = ReadBuffer(rel, insertPage);
 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
 
-		state = GenericXLogStart(index);
+		state = GenericXLogStart(rel);
 		page = GenericXLogRegisterBuffer(state, buf, 0);
 
 		if (PageGetFreeSpace(page) >= itemsz)
@@ -127,9 +124,9 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 			Page		newpage;
 
 			/* Add a new page */
-			LockRelationForExtension(index, ExclusiveLock);
+			LockRelationForExtension(rel, ExclusiveLock);
-			newbuf = IvfflatNewBuffer(index, MAIN_FORKNUM);
+			newbuf = IvfflatNewBuffer(rel, MAIN_FORKNUM);
-			UnlockRelationForExtension(index, ExclusiveLock);
+			UnlockRelationForExtension(rel, ExclusiveLock);
 
 			/* Init new page */
 			newpage = GenericXLogRegisterBuffer(state, newbuf, GENERIC_XLOG_FULL_IMAGE);
@@ -142,13 +139,15 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 			IvfflatPageGetOpaque(page)->nextblkno = insertPage;
 
 			/* Commit */
+			MarkBufferDirty(newbuf);
+			MarkBufferDirty(buf);
 			GenericXLogFinish(state);
 
 			/* Unlock previous buffer */
 			UnlockReleaseBuffer(buf);
 
 			/* Prepare new buffer */
-			state = GenericXLogStart(index);
+			state = GenericXLogStart(rel);
 			buf = newbuf;
 			page = GenericXLogRegisterBuffer(state, buf, 0);
 			break;
@@ -157,13 +156,13 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 
 	/* Add to next offset */
 	if (PageAddItem(page, (Item) itup, itemsz, InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
-		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
+		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(rel));
 
 	IvfflatCommitBuffer(buf, state);
 
 	/* Update the insert page */
 	if (insertPage != originalInsertPage)
-		IvfflatUpdateList(index, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
+		IvfflatUpdateList(rel, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
 }
 
 /*

src/ivfkmeans.c

@@ -6,10 +6,6 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 
-#ifdef IVFFLAT_MEMORY
-#include "utils/memutils.h"
-#endif
-
 /*
  * Initialize with kmeans++
  *
@@ -21,6 +17,10 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 	FmgrInfo   *procinfo;
 	Oid			collation;
 	int64		j;
+	double		distance;
+	double		sum;
+	double		choice;
+	Vector	   *vec;
 	float	   *weight = palloc(samples->length * sizeof(float));
 	int			numCenters = centers->maxlen;
 	int			numSamples = samples->length;
@@ -33,21 +33,17 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 	centers->length++;
 
 	for (j = 0; j < numSamples; j++)
-		weight[j] = FLT_MAX;
+		weight[j] = DBL_MAX;
 
 	for (int i = 0; i < numCenters; i++)
 	{
-		double		sum;
-		double		choice;
-
 		CHECK_FOR_INTERRUPTS();
 
 		sum = 0.0;
 
 		for (j = 0; j < numSamples; j++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, j);
+			vec = VectorArrayGet(samples, j);
-			double		distance;
 
 			/* Only need to compute distance for new center */
 			/* TODO Use triangle inequality to reduce distance calculations */
@@ -116,6 +112,7 @@ CompareVectors(const void *a, const void *b)
 static void
 QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 {
+	Vector	   *vec;
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
@@ -126,7 +123,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
 		for (int i = 0; i < samples->length; i++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, i);
+			vec = VectorArrayGet(samples, i);
 
 			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
 			{
@@ -139,7 +136,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 	/* Fill remaining with random data */
 	while (centers->length < centers->maxlen)
 	{
-		Vector	   *vec = VectorArrayGet(centers, centers->length);
+		vec = VectorArrayGet(centers, centers->length);
 
 		SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
 		vec->dim = dimensions;
@@ -155,23 +152,6 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 	}
 }
 
-#ifdef IVFFLAT_MEMORY
-/*
- * Show memory usage
- */
-static void
-ShowMemoryUsage(Size estimatedSize)
-{
-#if PG_VERSION_NUM >= 130000
-	elog(INFO, "total memory: %zu MB",
-		 MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
-#else
-	MemoryContextStats(CurrentMemoryContext);
-#endif
-	elog(INFO, "estimated memory: %zu MB", estimatedSize / (1024 * 1024));
-}
-#endif
-
 /*
  * Use Elkan for performance. This requires distance function to satisfy triangle inequality.
  *
@@ -188,6 +168,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	Oid			collation;
 	Vector	   *vec;
 	Vector	   *newCenter;
+	int			iteration;
 	int64		j;
 	int64		k;
 	int			dimensions = centers->dim;
@@ -201,6 +182,14 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	float	   *s;
 	float	   *halfcdist;
 	float	   *newcdist;
+	int			changes;
+	double		minDistance;
+	int			closestCenter;
+	double		distance;
+	bool		rj;
+	bool		rjreset;
+	double		dxcx;
+	double		dxc;
 
 	/* Calculate allocation sizes */
 	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
@@ -252,24 +241,20 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		vec->dim = dimensions;
 	}
 
-#ifdef IVFFLAT_MEMORY
-	ShowMemoryUsage(totalSize);
-#endif
-
 	/* Pick initial centers */
 	InitCenters(index, samples, centers, lowerBound);
 
 	/* Assign each x to its closest initial center c(x) = argmin d(x,c) */
 	for (j = 0; j < numSamples; j++)
 	{
-		float		minDistance = FLT_MAX;
+		minDistance = DBL_MAX;
-		int			closestCenter = 0;
+		closestCenter = 0;
 
 		/* Find closest center */
 		for (k = 0; k < numCenters; k++)
 		{
 			/* TODO Use Lemma 1 in k-means++ initialization */
-			float		distance = lowerBound[j * numCenters + k];
+			distance = lowerBound[j * numCenters + k];
 
 			if (distance < minDistance)
 			{
@@ -283,14 +268,13 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	}
 
 	/* Give 500 iterations to converge */
-	for (int iteration = 0; iteration < 500; iteration++)
+	for (iteration = 0; iteration < 500; iteration++)
 	{
-		int			changes = 0;
-		bool		rjreset;
-
 		/* Can take a while, so ensure we can interrupt */
 		CHECK_FOR_INTERRUPTS();
 
+		changes = 0;
+
 		/* Step 1: For all centers, compute distance */
 		for (j = 0; j < numCenters; j++)
 		{
@@ -298,8 +282,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 			for (k = j + 1; k < numCenters; k++)
 			{
-				float		distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
+				distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
-
 				halfcdist[j * numCenters + k] = distance;
 				halfcdist[k * numCenters + j] = distance;
 			}
@@ -308,12 +291,10 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		/* For all centers c, compute s(c) */
 		for (j = 0; j < numCenters; j++)
 		{
-			float		minDistance = FLT_MAX;
+			minDistance = DBL_MAX;
 
 			for (k = 0; k < numCenters; k++)
 			{
-				float		distance;
-
 				if (j == k)
 					continue;
 
@@ -329,8 +310,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 		for (j = 0; j < numSamples; j++)
 		{
-			bool		rj;
-
 			/* Step 2: Identify all points x such that u(x) <= s(c(x)) */
 			if (upperBound[j] <= s[closestCenters[j]])
 				continue;
@@ -339,8 +318,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 			for (k = 0; k < numCenters; k++)
 			{
-				float		dxcx;
-
 				/* Step 3: For all remaining points x and centers c */
 				if (k == closestCenters[j])
 					continue;
@@ -370,7 +347,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 				/* Step 3b */
 				if (dxcx > lowerBound[j * numCenters + k] || dxcx > halfcdist[closestCenters[j] * numCenters + k])
 				{
-					float		dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
+					dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
 
 					/* d(x,c) calculated */
 					lowerBound[j * numCenters + k] = dxc;
@@ -384,6 +361,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 						changes++;
 					}
+
 				}
 			}
 		}
@@ -400,8 +378,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 		for (j = 0; j < numSamples; j++)
 		{
-			int			closestCenter;
-
 			vec = VectorArrayGet(samples, j);
 			closestCenter = closestCenters[j];
 
@@ -450,7 +426,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		{
 			for (k = 0; k < numCenters; k++)
 			{
-				float		distance = lowerBound[j * numCenters + k] - newcdist[k];
+				distance = lowerBound[j * numCenters + k] - newcdist[k];
 
 				if (distance < 0)
 					distance = 0;
@@ -466,7 +442,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 
 		/* Step 7 */
 		for (j = 0; j < numCenters; j++)
-			VectorArraySet(centers, j, VectorArrayGet(newCenters, j));
+			memcpy(VectorArrayGet(centers, j), VectorArrayGet(newCenters, j), VECTOR_SIZE(dimensions));
 
 		if (changes == 0 && iteration != 0)
 			break;
@@ -489,6 +465,9 @@ static void
 CheckCenters(Relation index, VectorArray centers)
 {
 	FmgrInfo   *normprocinfo;
+	Oid			collation;
+	Vector	   *vec;
+	double		norm;
 
 	if (centers->length != centers->maxlen)
 		elog(ERROR, "Not enough centers. Please report a bug.");
@@ -496,7 +475,7 @@ CheckCenters(Relation index, VectorArray centers)
 	/* Ensure no NaN or infinite values */
 	for (int i = 0; i < centers->length; i++)
 	{
-		Vector	   *vec = VectorArrayGet(centers, i);
+		vec = VectorArrayGet(centers, i);
 
 		for (int j = 0; j < vec->dim; j++)
 		{
@@ -522,12 +501,11 @@ CheckCenters(Relation index, VectorArray centers)
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		Oid			collation = index->rd_indcollation[0];
+		collation = index->rd_indcollation[0];
 
 		for (int i = 0; i < centers->length; i++)
 		{
-			double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
+			norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
-
 			if (norm == 0)
 				elog(ERROR, "Zero norm detected. Please report a bug.");
 		}

src/ivfscan.c

@@ -31,36 +31,36 @@ CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
 static void
 GetScanLists(IndexScanDesc scan, Datum value)
 {
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	Buffer		cbuf;
+	Page		cpage;
+	IvfflatList list;
+	OffsetNumber offno;
+	OffsetNumber maxoffno;
 	BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
 	int			listCount = 0;
+	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
+	double		distance;
+	IvfflatScanList *scanlist;
 	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 (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
+			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));
 
 			if (listCount < so->probes)
 			{
-				IvfflatScanList *scanlist;
-
 				scanlist = &so->lists[listCount];
 				scanlist->startPage = list->startPage;
 				scanlist->distance = distance;
@@ -75,8 +75,6 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			}
 			else if (distance < maxDistance)
 			{
-				IvfflatScanList *scanlist;
-
 				/* Remove */
 				scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
 
@@ -103,6 +101,14 @@ 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;
 	TupleDesc	tupdesc = RelationGetDescr(scan->indexRelation);
 	double		tuples = 0;
 
@@ -122,28 +128,19 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	/* Search closest probes lists */
 	while (!pairingheap_is_empty(so->listQueue))
 	{
-		BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
+		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 (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+			for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 			{
-				IndexTuple	itup;
+				itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
-				Datum		datum;
-				bool		isnull;
-				ItemId		itemid = PageGetItemId(page, offno);
-
-				itup = (IndexTuple) PageGetItem(page, itemid);
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
 
 				/*
@@ -157,6 +154,8 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
+				slot->tts_values[2] = Int32GetDatum((int) searchPage);
+				slot->tts_isnull[2] = false;
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
@@ -181,6 +180,29 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
+/*
+ * Get dimensions from metapage
+ */
+static int
+GetDimensions(Relation index)
+{
+	Buffer		buf;
+	Page		page;
+	IvfflatMetaPage metap;
+	int			dimensions;
+
+	buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
+	LockBuffer(buf, BUFFER_LOCK_SHARE);
+	page = BufferGetPage(buf);
+	metap = IvfflatPageGetMeta(page);
+
+	dimensions = metap->dimensions;
+
+	UnlockReleaseBuffer(buf);
+
+	return dimensions;
+}
+
 /*
  * Prepare for an index scan
  */
@@ -190,7 +212,6 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	IndexScanDesc scan;
 	IvfflatScanOpaque so;
 	int			lists;
-	int			dimensions;
 	AttrNumber	attNums[] = {1};
 	Oid			sortOperators[] = {Float8LessOperator};
 	Oid			sortCollations[] = {InvalidOid};
@@ -198,31 +219,31 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	int			probes = ivfflat_probes;
 
 	scan = RelationGetIndexScan(index, nkeys, norderbys);
-
+	lists = IvfflatGetLists(scan->indexRelation);
-	/* 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->probes = probes;
-	so->dimensions = dimensions;
 
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	so->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
+	so->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	so->collation = index->rd_indcollation[0];
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(2);
+	so->tupdesc = CreateTemplateTupleDesc(3);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(2, false);
+	so->tupdesc = CreateTemplateTupleDesc(3, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
+	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -288,13 +309,8 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
-		/* Requires MVCC-compliant snapshot as not able to pin during sorting */
-		/* https://www.postgresql.org/docs/current/index-locking.html */
-		if (!IsMVCCSnapshot(scan->xs_snapshot))
-			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
-
 		if (scan->orderByData->sk_flags & SK_ISNULL)
-			value = PointerGetDatum(InitVector(so->dimensions));
+			value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
 		else
 		{
 			value = scan->orderByData->sk_argument;
@@ -304,8 +320,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 			Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
 
 			/* Fine if normalization fails */
-			if (so->normprocinfo != NULL)
+			IvfflatNormValue(so->normprocinfo, so->normalizeprocinfo, so->collation, &value, NULL);
-				IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL);
 		}
 
 		IvfflatBench("GetScanLists", GetScanLists(scan, value));
@@ -319,14 +334,26 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 
 	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
 	{
-		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
+		ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
+		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
 
 #if PG_VERSION_NUM >= 120000
-		scan->xs_heaptid = *heaptid;
+		scan->xs_heaptid = *tid;
 #else
-		scan->xs_ctup.t_self = *heaptid;
+		scan->xs_ctup.t_self = *tid;
 #endif
 
+		if (BufferIsValid(so->buf))
+			ReleaseBuffer(so->buf);
+
+		/*
+		 * An index scan must maintain a pin on the index page holding the
+		 * item last returned by amgettuple
+		 *
+		 * https://www.postgresql.org/docs/current/index-locking.html
+		 */
+		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
+
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -342,6 +369,10 @@ ivfflatendscan(IndexScanDesc scan)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 
+	/* Release pin */
+	if (BufferIsValid(so->buf))
+		ReleaseBuffer(so->buf);
+
 	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 

src/ivfutils.c

@@ -57,26 +57,35 @@ IvfflatGetLists(Relation index)
  * Get proc
  */
 FmgrInfo *
-IvfflatOptionalProcInfo(Relation index, uint16 procnum)
+IvfflatOptionalProcInfo(Relation rel, uint16 procnum)
 {
-	if (!OidIsValid(index_getprocid(index, 1, procnum)))
+	if (!OidIsValid(index_getprocid(rel, 1, procnum)))
 		return NULL;
 
-	return index_getprocinfo(index, 1, procnum);
+	return index_getprocinfo(rel, 1, procnum);
 }
 
 /*
- * Divide by the norm
+ * Normalize a vector
- *
- * Returns false if value should not be indexed
  *
  * The caller needs to free the pointer stored in value
  * if it's different than the original value
  */
-bool
+void
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
+IvfflatNormValue(FmgrInfo *procinfo, FmgrInfo *normalizeprocinfo, Oid collation, Datum *value, Vector * result)
 {
-	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
+	double		norm;
+
+	if (normalizeprocinfo != NULL)
+	{
+		*value = FunctionCall1Coll(normalizeprocinfo, collation, *value);
+		return;
+	}
+
+	if (procinfo == NULL)
+		return;
+
+	norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
 	{
@@ -89,11 +98,7 @@ IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * resul
 			result->x[i] = v->x[i] / norm;
 
 		*value = PointerGetDatum(result);
-
-		return true;
 	}
-
-	return false;
 }
 
 /*
@@ -136,6 +141,7 @@ IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogStat
 void
 IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
 {
+	MarkBufferDirty(buf);
 	GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 }
@@ -159,6 +165,8 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
 	IvfflatInitPage(newbuf, newpage);
 
 	/* Commit */
+	MarkBufferDirty(*buf);
+	MarkBufferDirty(newbuf);
 	GenericXLogFinish(*state);
 
 	/* Unlock */
@@ -169,29 +177,6 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
 	*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
  */

src/ivfvacuum.c

@@ -107,6 +107,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 				{
 					/* Delete tuples */
 					PageIndexMultiDelete(page, deletable, ndeletable);
+					MarkBufferDirty(buf);
 					GenericXLogFinish(state);
 				}
 				else

src/vector.c

@@ -15,10 +15,6 @@
 #include "utils/numeric.h"
 #include "vector.h"
 
-#if PG_VERSION_NUM >= 160000
-#include "varatt.h"
-#endif
-
 #if PG_VERSION_NUM >= 120000
 #include "common/shortest_dec.h"
 #include "utils/float.h"
@@ -39,7 +35,6 @@ PG_MODULE_MAGIC;
 /*
  * Initialize index options and variables
  */
-PGDLLEXPORT void _PG_init(void);
 void
 _PG_init(void)
 {
@@ -89,7 +84,7 @@ CheckDim(int dim)
 }
 
 /*
- * Ensure finite element
+ * Ensure finite elements
  */
 static inline void
 CheckElement(float value)
@@ -105,23 +100,6 @@ CheckElement(float value)
 				 errmsg("infinite value not allowed in vector")));
 }
 
-/*
- * Allocate and initialize a new vector
- */
-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;
-}
-
 /*
  * Check for whitespace, since array_isspace() is static
  */
@@ -177,15 +155,14 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_in);
 Datum
 vector_in(PG_FUNCTION_ARGS)
 {
-	char	   *lit = PG_GETARG_CSTRING(0);
+	char	   *str = PG_GETARG_CSTRING(0);
 	int32		typmod = PG_GETARG_INT32(2);
 	float		x[VECTOR_MAX_DIM];
 	int			dim = 0;
 	char	   *pt;
 	char	   *stringEnd;
 	Vector	   *result;
-	char	   *litcopy = pstrdup(lit);
+	char	   *lit = pstrdup(str);
-	char	   *str = litcopy;
 
 	while (vector_isspace(*str))
 		str++;
@@ -269,7 +246,7 @@ vector_in(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_DATA_EXCEPTION),
 				 errmsg("vector must have at least 1 dimension")));
 
-	pfree(litcopy);
+	pfree(lit);
 
 	CheckExpectedDim(typmod, dim);
 
@@ -438,18 +415,17 @@ vector_send(PG_FUNCTION_ARGS)
 
 /*
  * Convert vector to vector
- * This is needed to check the type modifier
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(vector);
 Datum
 vector(PG_FUNCTION_ARGS)
 {
-	Vector	   *vec = PG_GETARG_VECTOR_P(0);
+	Vector	   *arg = PG_GETARG_VECTOR_P(0);
 	int32		typmod = PG_GETARG_INT32(1);
 
-	CheckExpectedDim(typmod, vec->dim);
+	CheckExpectedDim(typmod, arg->dim);
 
-	PG_RETURN_POINTER(vec);
+	PG_RETURN_POINTER(arg);
 }
 
 /*
@@ -466,6 +442,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 	bool		typbyval;
 	char		typalign;
 	Datum	   *elemsp;
+	bool	   *nullsp;
 	int			nelemsp;
 
 	if (ARR_NDIM(array) > 1)
@@ -479,7 +456,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("array must not contain nulls")));
 
 	get_typlenbyvalalign(ARR_ELEMTYPE(array), &typlen, &typbyval, &typalign);
-	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, NULL, &nelemsp);
+	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, &nullsp, &nelemsp);
 
 	CheckDim(nelemsp);
 	CheckExpectedDim(typmod, nelemsp);
@@ -513,12 +490,6 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("unsupported array type")));
 	}
 
-	/*
-	 * Free allocation from deconstruct_array. Do not free individual elements
-	 * when pass-by-reference since they point to original array.
-	 */
-	pfree(elemsp);
-
 	/* Check elements */
 	for (int i = 0; i < result->dim; i++)
 		CheckElement(result->x[i]);
@@ -702,8 +673,6 @@ vector_spherical_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;
 	float		dp = 0.0;
 	double		distance;
 
@@ -711,7 +680,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 
 	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
-		dp += ax[i] * bx[i];
+		dp += a->x[i] * b->x[i];
 
 	distance = (double) dp;
 
@@ -725,7 +694,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 }
 
 /*
- * Get the L1 distance between two vectors
+ * Get the L1 distance between vectors
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(l1_distance);
 Datum
@@ -764,16 +733,62 @@ vector_dims(PG_FUNCTION_ARGS)
 PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_norm);
 Datum
 vector_norm(PG_FUNCTION_ARGS)
+{
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	float	   *ax = a->x;
+	float		norm = 0.0;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		norm += ax[i] * ax[i];
+
+	PG_RETURN_FLOAT8(sqrt((double) norm));
+}
+
+/*
+ * Normalize a vector with the L2 norm
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(normalize_l2);
+Datum
+normalize_l2(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	float	   *ax = a->x;
 	double		norm = 0.0;
+#ifdef _MSC_VER
+	/* Fix precision on Windows */
+	double		normf;
+#else
+	float		normf;
+#endif
+	Vector	   *result;
+	float	   *rx;
+
+	result = InitVector(a->dim);
+	rx = result->x;
 
 	/* Auto-vectorized */
 	for (int i = 0; i < a->dim; i++)
 		norm += (double) ax[i] * (double) ax[i];
 
-	PG_RETURN_FLOAT8(sqrt(norm));
+	norm = sqrt(norm);
+	normf = norm;
+
+	if (normf > 0)
+	{
+		/* Auto-vectorized */
+		for (int i = 0, imax = a->dim; i < imax; i++)
+			rx[i] = ax[i] / normf;
+
+		/* Check for overflow */
+		for (int i = 0, imax = a->dim; i < imax; i++)
+		{
+			if (isinf(rx[i]))
+				float_overflow_error();
+		}
+	}
+
+	PG_RETURN_POINTER(result);
 }
 
 /*
@@ -904,8 +919,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_lt);
 Datum
 vector_lt(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) < 0);
 }
@@ -917,8 +932,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_le);
 Datum
 vector_le(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) <= 0);
 }
@@ -930,8 +945,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_eq);
 Datum
 vector_eq(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) == 0);
 }
@@ -943,8 +958,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ne);
 Datum
 vector_ne(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) != 0);
 }
@@ -956,8 +971,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ge);
 Datum
 vector_ge(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) >= 0);
 }
@@ -969,8 +984,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_gt);
 Datum
 vector_gt(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) > 0);
 }
@@ -982,8 +997,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_cmp);
 Datum
 vector_cmp(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
-	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_INT32(vector_cmp_internal(a, b));
 }

src/vector.h

@@ -1,6 +1,12 @@
 #ifndef VECTOR_H
 #define VECTOR_H
 
+#include "postgres.h"
+
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
 #define VECTOR_MAX_DIM 16000
 
 #define VECTOR_SIZE(_dim)		(offsetof(Vector, x) + sizeof(float)*(_dim))
@@ -8,6 +14,9 @@
 #define PG_GETARG_VECTOR_P(x)	DatumGetVector(PG_GETARG_DATUM(x))
 #define PG_RETURN_VECTOR_P(x)	PG_RETURN_POINTER(x)
 
+/* Exported functions */
+PGDLLEXPORT void _PG_init(void);
+
 typedef struct Vector
 {
 	int32		vl_len_;		/* varlena header (do not touch directly!) */
@@ -16,8 +25,24 @@ 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/functions.out

@@ -48,10 +48,34 @@ SELECT vector_norm('[0,1]');
            1
 (1 row)
 
-SELECT vector_norm('[3e37,4e37]')::real;
+SELECT normalize_l2('[3,4]');
- vector_norm 
+ normalize_l2 
--------------
+--------------
-       5e+37
+ [0.6,0.8]
+(1 row)
+
+SELECT normalize_l2('[3,0]');
+ normalize_l2 
+--------------
+ [1,0]
+(1 row)
+
+SELECT normalize_l2('[0,0.1]');
+ normalize_l2 
+--------------
+ [0,1]
+(1 row)
+
+SELECT normalize_l2('[0,0]');
+ normalize_l2 
+--------------
+ [0,0]
+(1 row)
+
+SELECT normalize_l2('[3e38]');
+ normalize_l2 
+--------------
+ [1]
 (1 row)
 
 SELECT l2_distance('[0,0]', '[3,4]');
@@ -106,12 +130,6 @@ SELECT cosine_distance('[1,1]', '[1,1]');
                0
 (1 row)
 
-SELECT cosine_distance('[1,0]', '[0,2]');
- cosine_distance 
------------------
-               1
-(1 row)
-
 SELECT cosine_distance('[1,1]', '[-1,-1]');
  cosine_distance 
 -----------------

test/expected/hnsw_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/hnsw_l2.out

@@ -27,10 +27,4 @@ SELECT COUNT(*) FROM t;
      5
 (1 row)
 
-TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
- val 
------
-(0 rows)
-
 DROP TABLE t;

test/expected/hnsw_options.out

@@ -1,26 +1,25 @@
+SET enable_seqscan = off;
 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 = 3);
-ERROR:  value 1 out of bounds for option "m"
+ERROR:  value 3 out of bounds for option "m"
-DETAIL:  Valid values are between "2" and "100".
+DETAIL:  Valid values are between "4" 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".
+DETAIL:  Valid values are between "4" and "100".
-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 = 9);
-ERROR:  value 3 out of bounds for option "ef_construction"
+ERROR:  value 9 out of bounds for option "ef_construction"
-DETAIL:  Valid values are between "4" and "1000".
+DETAIL:  Valid values are between "10" 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".
+DETAIL:  Valid values are between "10" 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;
+SET hnsw.ef_search = 9;
-ERROR:  0 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
+ERROR:  9 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
 SET hnsw.ef_search = 1001;
-ERROR:  1001 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
+ERROR:  1001 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
 DROP TABLE t;

test/expected/ivfflat_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/ivfflat_l2.out

@@ -1,7 +1,7 @@
 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);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
    val   
@@ -27,13 +27,4 @@ SELECT COUNT(*) FROM t;
      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

@@ -1,8 +1,9 @@
+SET enable_seqscan = off;
 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) 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);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
 ERROR:  value 32769 out of bounds for option "lists"
 DETAIL:  Valid values are between "1" and "32768".
 SHOW ivfflat.probes;

test/expected/ivfflat_unlogged.out

@@ -1,7 +1,7 @@
 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 vector_l2_ops) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
    val   
 ---------

test/sql/functions.sql

@@ -11,7 +11,12 @@ 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 normalize_l2('[3,4]');
+SELECT normalize_l2('[3,0]');
+SELECT normalize_l2('[0,0.1]');
+SELECT normalize_l2('[0,0]');
+SELECT normalize_l2('[3e38]');
 
 SELECT l2_distance('[0,0]', '[3,4]');
 SELECT l2_distance('[0,0]', '[0,1]');
@@ -25,7 +30,6 @@ 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]');

test/sql/hnsw_l2.sql

@@ -10,7 +10,4 @@ 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

@@ -1,13 +1,14 @@
+SET enable_seqscan = off;
+
 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 = 3);
 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 = 9);
 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 = 9;
 SET hnsw.ef_search = 1001;
 
 DROP TABLE t;

test/sql/ivfflat_l2.sql

@@ -2,7 +2,7 @@ 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);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
@@ -10,7 +10,4 @@ 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

@@ -1,6 +1,8 @@
+SET enable_seqscan = off;
+
 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) WITH (lists = 0);
-CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
 
 SHOW ivfflat.probes;
 

test/sql/ivfflat_unlogged.sql

@@ -2,7 +2,7 @@ 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 vector_l2_ops) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
 
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
 

test/t/001_ivfflat_wal.pl

@@ -19,13 +19,14 @@ sub test_index_replay
 
 	# Wait for replica to catch up
 	my $applname = $node_replica->name;
+
+	my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
 	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 @r = ();
-	for (1 .. $dim)
+	for (1 .. $dim) {
-	{
 		push(@r, rand());
 	}
 	my $sql = join(",", @r);
@@ -51,13 +52,11 @@ 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)
+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)
+if ($dim > 1500) {
-{
 	$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
 }
 $node_primary->start;
@@ -68,7 +67,8 @@ $node_primary->backup($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
@@ -77,7 +77,7 @@ $node_primary->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));
 $node_primary->safe_psql("postgres",
 	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
-$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
 
 # Test that queries give same result
 test_index_replay('initial');

test/t/002_ivfflat_vacuum.pl

@@ -7,8 +7,7 @@ use Test::More;
 my $dim = 3;
 
 my @r = ();
-for (1 .. $dim)
+for (1 .. $dim) {
-{
 	my $v = int(rand(1000)) + 1;
 	push(@r, "i % $v");
 }
@@ -25,7 +24,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node->safe_psql("postgres",
 	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
 );
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
 
 # Get size
 my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");

test/t/003_ivfflat_recall.pl

@@ -22,8 +22,7 @@ sub test_recall
 	));
 	like($explain, qr/Index Scan using idx on tst/);
 
-	for my $i (0 .. $#queries)
+	for my $i (0 .. $#queries) {
-	{
 		my $actual = $node->safe_psql("postgres", qq(
 			SET enable_seqscan = off;
 			SET ivfflat.probes = $probes;
@@ -34,10 +33,8 @@ sub test_recall
 
 		my @expected_ids = split("\n", $expected[$i]);
 
-		foreach (@expected_ids)
+		foreach (@expected_ids) {
-		{
+			if (exists($actual_set{$_})) {
-			if (exists($actual_set{$_}))
-			{
 				$correct++;
 			}
 			$total++;
@@ -60,8 +57,7 @@ $node->safe_psql("postgres",
 );
 
 # Generate queries
-for (1 .. 20)
+for (1..20) {
-{
 	my $r1 = rand();
 	my $r2 = rand();
 	my $r3 = rand();
@@ -70,21 +66,26 @@ for (1 .. 20)
 
 # Check each index type
 my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
 
-for my $i (0 .. $#operators)
+foreach (@operators) {
-{
+	my $operator = $_;
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
 
 	# Get exact results
 	@expected = ();
-	foreach (@queries)
+	foreach (@queries) {
-	{
 		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
 		push(@expected, $res);
 	}
 
+	my $opclass;
+	if ($operator eq "<->") {
+		$opclass = "vector_l2_ops";
+	} elsif ($operator eq "<#>") {
+		$opclass = "vector_ip_ops";
+	} else {
+		$opclass = "vector_cosine_ops";
+	}
+
 	# Build index serially
 	$node->safe_psql("postgres", qq(
 		SET max_parallel_maintenance_workers = 0;
@@ -92,14 +93,13 @@ for my $i (0 .. $#operators)
 	));
 
 	# Test approximate results
-	if ($operator ne "<#>")
+	if ($operator ne "<#>") {
-	{
+		# TODO fix test
-		# TODO Fix test (uniform random vectors all have similar inner product)
+		test_recall(1, 0.75, $operator);
-		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}
 	# Account for equal distances
-	test_recall(100, 0.9925, $operator);
+	test_recall(100, 0.995, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");
 
@@ -113,14 +113,13 @@ for my $i (0 .. $#operators)
 	like($stderr, qr/using \d+ parallel workers/);
 
 	# Test approximate results
-	if ($operator ne "<#>")
+	if ($operator ne "<#>") {
-	{
+		# TODO fix test
-		# TODO Fix test (uniform random vectors all have similar inner product)
+		test_recall(1, 0.75, $operator);
-		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}
 	# Account for equal distances
-	test_recall(100, 0.9925, $operator);
+	test_recall(100, 0.995, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");
 }

test/t/004_ivfflat_centers.pl

@@ -20,7 +20,7 @@ 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 = $lists);");
+	my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v) WITH (lists = $lists);");
 	is($ret, 0, $stderr);
 }
 

test/t/005_ivfflat_query_recall.pl

@@ -18,21 +18,24 @@ $node->safe_psql("postgres",
 
 # Check each index type
 my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+foreach (@operators) {
-
+	my $operator = $_;
-for my $i (0 .. $#operators)
-{
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
 
 	# Add index
+	my $opclass;
+	if ($operator eq "<->") {
+		$opclass = "vector_l2_ops";
+	} elsif ($operator eq "<#>") {
+		$opclass = "vector_ip_ops";
+	} else {
+		$opclass = "vector_cosine_ops";
+	}
 	$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v $opclass);");
 
 	# Test 100% recall
-	for (1 .. 20)
+	for (1..20) {
-	{
+		my $i = int(rand() * 100000);
-		my $id = int(rand() * 100000);
+		my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $i;");
-		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;

test/t/006_ivfflat_lists.pl

@@ -16,8 +16,8 @@ $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 lists50 ON tst USING ivfflat (v) WITH (lists = 50);");
-$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 100);");
+$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v) WITH (lists = 100);");
 
 # Test prefers more lists
 my $res = $node->safe_psql("postgres", "EXPLAIN SELECT v FROM tst ORDER BY v <-> '[0.5,0.5,0.5]' LIMIT 10;");
@@ -26,7 +26,7 @@ unlike($res, qr/lists50/);
 
 # Test errors with too much memory
 my ($ret, $stdout, $stderr) = $node->psql("postgres",
-	"CREATE INDEX lists10000 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10000);"
+	"CREATE INDEX lists10000 ON tst USING ivfflat (v) WITH (lists = 10000);"
 );
 like($stderr, qr/memory required is/);
 

test/t/007_ivfflat_inserts.pl

@@ -19,7 +19,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (v vector($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->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
 
 $node->pgbench(
 	"--no-vacuum --client=5 --transactions=100",
@@ -28,7 +28,7 @@ $node->pgbench(
 	[qr{^$}],
 	"concurrent INSERTs",
 	{
-		"007_ivfflat_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
+		"007_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
 	}
 );
 

test/t/008_aggregates.pl

@@ -30,8 +30,7 @@ sub test_aggregate
 
 	# Test matches real for avg
 	# Cannot test sum since sum(real) varies between calls
-	if ($agg eq 'avg')
+	if ($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;");

test/t/010_hnsw_wal.pl

@@ -19,13 +19,14 @@ sub test_index_replay
 
 	# Wait for replica to catch up
 	my $applname = $node_replica->name;
+
+	my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
 	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 @r = ();
-	for (1 .. $dim)
+	for (1 .. $dim) {
-	{
 		push(@r, rand());
 	}
 	my $sql = join(",", @r);
@@ -51,13 +52,11 @@ 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)
+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)
+if ($dim > 1500) {
-{
 	$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
 }
 $node_primary->start;
@@ -68,7 +67,8 @@ $node_primary->backup($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 hnsw index on primary

test/t/011_hnsw_vacuum.pl

@@ -7,8 +7,7 @@ use Test::More;
 my $dim = 3;
 
 my @r = ();
-for (1 .. $dim)
+for (1 .. $dim) {
-{
 	my $v = int(rand(1000)) + 1;
 	push(@r, "i % $v");
 }
@@ -23,7 +22,7 @@ $node->start;
 $node->safe_psql("postgres", "CREATE EXTENSION vector;");
 $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
 );
 $node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
 
@@ -34,21 +33,11 @@ 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;");
 $node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+	"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
 );
 
 # Check size
-# May increase some due to different levels
 my $new_size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");
-cmp_ok($new_size, "<=", $size * 1.02, "size does not increase too much");
+cmp_ok($new_size, "<=", $size * 1.01, "size does not increase too much");
-
-# Delete all but one
-$node->safe_psql("postgres", "DELETE FROM tst WHERE i != 123;");
-$node->safe_psql("postgres", "VACUUM tst;");
-my $res = $node->safe_psql("postgres", qq(
-	SET enable_seqscan = off;
-	SELECT i FROM tst ORDER BY v <-> '[0,0,0]' LIMIT 10;
-));
-is($res, 123);
 
 done_testing();

test/t/012_hnsw_build_recall.pl

@@ -21,8 +21,7 @@ sub test_recall
 	));
 	like($explain, qr/Index Scan/);
 
-	for my $i (0 .. $#queries)
+	for my $i (0 .. $#queries) {
-	{
 		my $actual = $node->safe_psql("postgres", qq(
 			SET enable_seqscan = off;
 			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
@@ -32,10 +31,8 @@ sub test_recall
 
 		my @expected_ids = split("\n", $expected[$i]);
 
-		foreach (@expected_ids)
+		foreach (@expected_ids) {
-		{
+			if (exists($actual_set{$_})) {
-			if (exists($actual_set{$_}))
-			{
 				$correct++;
 			}
 			$total++;
@@ -58,8 +55,7 @@ $node->safe_psql("postgres",
 );
 
 # Generate queries
-for (1 .. 20)
+for (1..20) {
-{
 	my $r1 = rand();
 	my $r2 = rand();
 	my $r3 = rand();
@@ -68,47 +64,33 @@ for (1 .. 20)
 
 # Check each index type
 my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
 
-for my $i (0 .. $#operators)
+foreach (@operators) {
-{
+	my $operator = $_;
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
 
 	# Get exact results
 	@expected = ();
-	foreach (@queries)
+	foreach (@queries) {
-	{
 		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
 		push(@expected, $res);
 	}
 
-	# Build index serially
+	# Add index
-	$node->safe_psql("postgres", qq(
+	my $opclass;
-		SET max_parallel_maintenance_workers = 0;
+	if ($operator eq "<->") {
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+		$opclass = "vector_l2_ops";
-	));
+	} elsif ($operator eq "<#>") {
+		$opclass = "vector_ip_ops";
+	} else {
+		$opclass = "vector_cosine_ops";
+	}
+	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
 
-	# Test approximate results
+	if ($operator eq "<#>") {
-	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+		test_recall(0.80, $operator);
-	test_recall($min, $operator);
+	} else {
-
+		test_recall(0.99, $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;
-		SET hnsw.enable_parallel_build = on;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
-	is($ret, 0, $stderr);
-	like($stderr, qr/using \d+ parallel workers/);
-
-	# Test approximate results
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
 }
 
 done_testing();

test/t/013_hnsw_insert_recall.pl

@@ -21,8 +21,7 @@ sub test_recall
 	));
 	like($explain, qr/Index Scan/);
 
-	for my $i (0 .. $#queries)
+	for my $i (0 .. $#queries) {
-	{
 		my $actual = $node->safe_psql("postgres", qq(
 			SET enable_seqscan = off;
 			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
@@ -32,10 +31,8 @@ sub test_recall
 
 		my @expected_ids = split("\n", $expected[$i]);
 
-		foreach (@expected_ids)
+		foreach (@expected_ids) {
-		{
+			if (exists($actual_set{$_})) {
-			if (exists($actual_set{$_}))
-			{
 				$correct++;
 			}
 			$total++;
@@ -52,11 +49,10 @@ $node->start;
 
 # Create table
 $node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector(3));");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector(3));");
 
 # Generate queries
-for (1 .. 20)
+for (1..20) {
-{
 	my $r1 = rand();
 	my $r2 = rand();
 	my $r3 = rand();
@@ -65,32 +61,28 @@ for (1 .. 20)
 
 # Check each index type
 my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
 
-for my $i (0 .. $#operators)
+foreach (@operators) {
-{
+	my $operator = $_;
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
 
 	# Add index
+	my $opclass;
+	if ($operator eq "<->") {
+		$opclass = "vector_l2_ops";
+	} elsif ($operator eq "<#>") {
+		$opclass = "vector_ip_ops";
+	} else {
+		$opclass = "vector_cosine_ops";
+	}
 	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
 
-	# Use concurrent inserts
+	$node->safe_psql("postgres",
-	$node->pgbench(
+		"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
-		"--no-vacuum --client=10 --transactions=1000",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"013_hnsw_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES (ARRAY[random(), random(), random()]);"
-		}
 	);
 
 	# Get exact results
 	@expected = ();
-	foreach (@queries)
+	foreach (@queries) {
-	{
 		my $res = $node->safe_psql("postgres", qq(
 			SET enable_indexscan = off;
 			SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
@@ -98,8 +90,11 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+	if ($operator eq "<#>") {
-	test_recall($min, $operator);
+		test_recall(0.80, $operator);
+	} else {
+		test_recall(0.99, $operator);
+	}
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");
 	$node->safe_psql("postgres", "TRUNCATE tst;");

test/t/014_hnsw_inserts.pl

@@ -17,8 +17,22 @@ $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, 100) i;"
+);
 $node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
 
+$node->pgbench(
+	"--no-vacuum --client=5 --transactions=100",
+	0,
+	[qr{actually processed}],
+	[qr{^$}],
+	"concurrent INSERTs",
+	{
+		"007_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
+	}
+);
+
 sub idx_scan
 {
 	# Stats do not update instantaneously
@@ -27,48 +41,18 @@ sub idx_scan
 	$node->safe_psql("postgres", "SELECT idx_scan FROM pg_stat_user_indexes WHERE indexrelid = 'tst_v_idx'::regclass;");
 }
 
-for my $i (1 .. 20)
+my $expected = 100 + 5 * 100 * 10;
-{
-	$node->pgbench(
-		"--no-vacuum --client=10 --transactions=1",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"014_hnsw_inserts_$i" => "INSERT INTO tst VALUES (ARRAY[$array_sql]);"
-		}
-	);
 
-	my $count = $node->safe_psql("postgres", qq(
+my $count = $node->safe_psql("postgres", "SELECT COUNT(*) FROM tst;");
-		SET enable_seqscan = off;
+is($count, $expected);
-		SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
+is(idx_scan(), 0);
-	));
-	is($count, 10);
 
-	$node->safe_psql("postgres", "TRUNCATE tst;");
+$count = $node->safe_psql("postgres", qq(
-}
-
-$node->pgbench(
-	"--no-vacuum --client=20 --transactions=5",
-	0,
-	[qr{actually processed}],
-	[qr{^$}],
-	"concurrent INSERTs",
-	{
-		"014_hnsw_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
-	}
-);
-
-my $count = $node->safe_psql("postgres", qq(
 	SET enable_seqscan = off;
-	SET hnsw.ef_search = 1000;
+	SET hnsw.ef_search = 400;
 	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
 ));
-# Elements may lose all incoming connections with the HNSW algorithm
+is($count, 400);
-# Vacuuming can fix this if one of the elements neighbors is deleted
+is(idx_scan(), 1);
-cmp_ok($count, ">=", 997);
-
-is(idx_scan(), 21);
 
 done_testing();

test/t/015_hnsw_duplicates.pl

@@ -1,58 +0,0 @@
-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));");
-
-sub insert_vectors
-{
-	for my $i (1 .. 20)
-	{
-		$node->safe_psql("postgres", "INSERT INTO tst VALUES ('[1,1,1]');");
-	}
-}
-
-sub test_duplicates
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 1;
-		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]') t;
-	));
-	is($res, 10);
-}
-
-# Test duplicates with build
-insert_vectors();
-$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v vector_l2_ops);");
-test_duplicates();
-
-# Reset
-$node->safe_psql("postgres", "TRUNCATE tst;");
-
-# Test duplicates with inserts
-insert_vectors();
-test_duplicates();
-
-# Test fallback path for inserts
-$node->pgbench(
-	"--no-vacuum --client=5 --transactions=100",
-	0,
-	[qr{actually processed}],
-	[qr{^$}],
-	"concurrent INSERTs",
-	{
-		"015_hnsw_duplicates" => "INSERT INTO tst VALUES ('[1,1,1]');"
-	}
-);
-
-done_testing();

test/t/016_hnsw_vacuum_recall.pl

@@ -1,97 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-
-sub test_recall
-{
-	my ($min, $ef_search, $test_name) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = $ef_search;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$queries[0]' LIMIT $limit;
-	));
-	like($explain, qr/Index Scan/);
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SET hnsw.ef_search = $ef_search;
-			SELECT i FROM tst ORDER BY v <-> '$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, $test_name);
-}
-
-# Initialize node
-$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));");
-$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
-);
-
-# Add index
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops) WITH (m = 4, ef_construction = 8);");
-
-# Delete data
-$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r1 = rand();
-	my $r2 = rand();
-	my $r3 = rand();
-	push(@queries, "[$r1,$r2,$r3]");
-}
-
-# Get exact results
-@expected = ();
-foreach (@queries)
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_indexscan = off;
-		SELECT i FROM tst ORDER BY v <-> '$_' LIMIT $limit;
-	));
-	push(@expected, $res);
-}
-
-test_recall(0.20, $limit, "before vacuum");
-test_recall(0.95, 100, "before vacuum");
-
-# TODO Test concurrent inserts with vacuum
-$node->safe_psql("postgres", "VACUUM tst;");
-
-test_recall(0.95, $limit, "after vacuum");
-
-done_testing();

test/t/017_ivfflat_insert_recall.pl

@@ -1,117 +0,0 @@
-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 serial, v vector(3));");
-
-# 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];
-
-	# Add index
-	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING ivfflat (v $opclass);");
-
-	# Use concurrent inserts
-	$node->pgbench(
-		"--no-vacuum --client=10 --transactions=1000",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"017_ivfflat_insert_recall_$opclass" => "INSERT INTO tst (v) SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 10) i;"
-		}
-	);
-
-	# Get exact results
-	@expected = ();
-	foreach (@queries)
-	{
-		my $res = $node->safe_psql("postgres", qq(
-			SET enable_indexscan = off;
-			SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
-		));
-		push(@expected, $res);
-	}
-
-	# Test approximate results
-	if ($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;");
-	$node->safe_psql("postgres", "TRUNCATE tst;");
-}
-
-done_testing();

test/t/019_ivfflat_limit.pl

@@ -1,64 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-# Initialize node
-my $node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table and index
-$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, 1000) i;"
-);
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10);");
-
-# Test limit
-my $explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 100;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit with probes
-$explain = $node->safe_psql("postgres", qq(
-	SET ivfflat.probes = 2;
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 200;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit + offset
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 90 OFFSET 10;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit > expected tuples
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 101;
-));
-like($explain, qr/Seq Scan/);
-
-# Test limit > expected tuples with probes
-$explain = $node->safe_psql("postgres", qq(
-	SET ivfflat.probes = 2;
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 201;
-));
-like($explain, qr/Seq Scan/);
-
-# Test limit + offset > expected tuples
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 91 OFFSET 10;
-));
-like($explain, qr/Seq Scan/);
-
-# Test no limit
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]';
-));
-like($explain, qr/Seq Scan/);
-
-done_testing();

test/t/020_hnsw_limit.pl

@@ -1,62 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-# Initialize node
-my $node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table and index
-$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, 1000) i;"
-);
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
-
-# Test limit
-my $explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 40;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit with CTE
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE WITH cte AS (SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 40) SELECT * FROM cte;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit + offset
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 30 OFFSET 10;
-));
-like($explain, qr/Index Scan/);
-
-# Test limit > ef_search
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 41;
-));
-like($explain, qr/Seq Scan/);
-
-# Test limit > ef_search with CTE
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE WITH cte AS (SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 41) SELECT * FROM cte;
-));
-like($explain, qr/Seq Scan/);
-
-# Test limit + offset > ef_search
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]' LIMIT 31 OFFSET 10;
-));
-like($explain, qr/Seq Scan/);
-
-# Test no limit
-$explain = $node->safe_psql("postgres", qq(
-	EXPLAIN ANALYZE SELECT * FROM tst ORDER BY v <-> '[1,2,3]';
-));
-like($explain, qr/Seq Scan/);
-
-done_testing();

vector.control

@@ -1,4 +1,4 @@
-comment = 'vector data type and ivfflat and hnsw access methods'
+comment = 'vector data type and ivfflat access method'
-default_version = '0.5.1'
+default_version = '0.4.4'
 module_pathname = '$libdir/vector'
 relocatable = true