Allow override

This reverts commit 4894dc5da1.

.github/workflows/build.yml

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

CHANGELOG.md

@@ -1,3 +1,12 @@
+## 0.5.2 (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

Makefile.win

@@ -56,7 +56,7 @@ install:
 	copy $(EXTENSION).control "$(SHAREDIR)\extension"
 	copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
 	mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
-	copy $(HEADERS) "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+	for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 installcheck:
 	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)

README.md

@@ -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/workflows/build/badge.svg?branch=master)](https://github.com/pgvector/pgvector/actions)
+[![Build Status](https://github.com/pgvector/pgvector/actions/workflows/build.yml/badge.svg)](https://github.com/pgvector/pgvector/actions)
 
 ## Installation
 
@@ -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)
+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).
 
 ## Getting Started
 
@@ -161,8 +161,97 @@ You can add an index to use approximate nearest neighbor search, which trades so
 
 Supported index types are:
 
-- [IVFFlat](#ivfflat)
 - [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
 
@@ -215,78 +304,32 @@ SELECT ...
 COMMIT;
 ```
 
-## HNSW
+### Index Build Time
 
-An HNSW index creates a multilayer graph. It has slower build times and uses more memory than IVFFlat, but has better query performance (in terms of speed-recall tradeoff). There’s no training step like IVFFlat, so the index can be created without any data in the table.
+Speed up index creation on large tables by increasing the number of parallel workers (2 by default)
-
-Add an index for each distance function you want to use.
-
-L2 distance
 
 ```sql
-CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
+SET max_parallel_maintenance_workers = 7; -- plus leader
 ```
 
-Inner product
+For a large number of workers, you may also need to increase `max_parallel_workers` (8 by default)
 
-```sql
+### Indexing Progress
-CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
-```
-
-Cosine distance
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
-```
-
-Vectors with up to 2,000 dimensions can be indexed.
-
-### Index Options
-
-Specify HNSW parameters
-
-- `m` - the max number of connections per layer (16 by default)
-- `ef_construction` - the size of the dynamic candidate list for constructing the graph (64 by default)
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WITH (m = 16, ef_construction = 64);
-```
-
-### Query Options
-
-Specify the size of the dynamic candidate list for search (40 by default)
-
-```sql
-SET hnsw.ef_search = 100;
-```
-
-A higher value provides better recall at the cost of speed.
-
-Use `SET LOCAL` inside a transaction to set it for a single query
-
-```sql
-BEGIN;
-SET LOCAL hnsw.ef_search = 100;
-SELECT ...
-COMMIT;
-```
-
-## Indexing Progress
 
 Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
 
 ```sql
-SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
+SELECT phase, round(100.0 * tuples_done / nullif(tuples_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
 ```
 
-The phases are:
+The phases for IVFFlat are:
 
 1. `initializing`
-2. `performing k-means` - IVFFlat only
+2. `performing k-means`
-3. `assigning tuples` - IVFFlat only
+3. `assigning tuples`
 4. `loading tuples`
 
-Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
+Note: `%` is only populated during the `loading tuples` phase
 
 ## Filtering
 
@@ -305,8 +348,7 @@ CREATE INDEX ON items (category_id);
 Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search
 
 ```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WHERE (category_id = 123);
-    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
@@ -317,13 +359,15 @@ 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 ([Python example](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py)).
+Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search.
 
 ```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.
@@ -360,17 +404,21 @@ 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# | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
+C#, F#, Visual Basic | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
 Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
 Dart | [pgvector-dart](https://github.com/pgvector/pgvector-dart)
 Elixir | [pgvector-elixir](https://github.com/pgvector/pgvector-elixir)
 Go | [pgvector-go](https://github.com/pgvector/pgvector-go)
 Haskell | [pgvector-haskell](https://github.com/pgvector/pgvector-haskell)
-Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
+Java, Kotlin, Groovy, 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)
-Node.js | [pgvector-node](https://github.com/pgvector/pgvector-node)
+Nim | [pgvector-nim](https://github.com/pgvector/pgvector-nim)
+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)
@@ -378,6 +426,7 @@ 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
 
@@ -393,6 +442,63 @@ 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?
@@ -406,6 +512,8 @@ 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:
@@ -474,7 +582,7 @@ 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=/Applications/Postgres.app/Contents/Versions/latest/bin/pg_config
+export PG_CONFIG=/Library/PostgreSQL/16/bin/pg_config
 ```
 
 Then re-run the installation instructions (run `make clean` before `make` if needed). If `sudo` is needed for `make install`, use:
@@ -483,6 +591,14 @@ 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.
@@ -490,10 +606,14 @@ 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-15
+sudo apt install postgresql-server-dev-16
 ```
 
-Note: Replace `15` with your Postgres server version
+Note: Replace `16` 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
 
@@ -508,7 +628,7 @@ 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\15"
+set "PGROOT=C:\Program Files\PostgreSQL\16"
 git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
@@ -558,22 +678,22 @@ 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-15-pgvector
+sudo apt install postgresql-16-pgvector
 ```
 
-Note: Replace `15` with your Postgres server version
+Note: Replace `16` with your Postgres server version
 
 ### Yum
 
 RPM packages are available from the [PostgreSQL Yum Repository](https://yum.postgresql.org/). Follow the [setup instructions](https://www.postgresql.org/download/linux/redhat/) for your distribution and run:
 
 ```sh
-sudo yum install pgvector_15
+sudo yum install pgvector_16
 # or
-sudo dnf install pgvector_15
+sudo dnf install pgvector_16
 ```
 
-Note: Replace `15` with your Postgres server version
+Note: Replace `16` with your Postgres server version
 
 ### conda-forge
 
@@ -595,7 +715,7 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
 
 ## Upgrading
 
-Install the latest version. Then in each database you want to upgrade, run:
+[Install](#installation) the latest version (use the same method as the original installation). Then in each database you want to upgrade, run:
 
 ```sql
 ALTER EXTENSION vector UPDATE;
@@ -683,7 +803,25 @@ 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

src/hnsw.c

@@ -14,6 +14,7 @@
 #endif
 
 int			hnsw_ef_search;
+bool		hnsw_enable_parallel_build;
 static relopt_kind hnsw_relopt_kind;
 
 /*
@@ -39,6 +40,11 @@ HnswInit(void)
 	DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
 							"Valid range is 1..1000.", &hnsw_ef_search,
 							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
+
+	/* 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);
 }
 
 /*
@@ -77,8 +83,8 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	List	   *qinfos;
 #endif
 
-	/* Never use index without order */
+	/* Never use index without order or limit */
-	if (path->indexorderbys == NULL)
+	if (path->indexorderbys == NULL || root->limit_tuples < 0)
 	{
 		*indexStartupCost = DBL_MAX;
 		*indexTotalCost = DBL_MAX;
@@ -88,6 +94,20 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 		return;
 	}
 
+	/*
+	 * Do not use index if limit + offset > ef_search unless enable_seqscan =
+	 * off
+	 */
+	if (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);

src/hnsw.h

@@ -4,7 +4,9 @@
 #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"
@@ -14,6 +16,10 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
+#if PG_VERSION_NUM < 120000
+#include "access/relscan.h"
+#endif
+
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
@@ -59,7 +65,7 @@
 
 #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))
@@ -67,8 +73,10 @@
 
 #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
@@ -86,24 +94,28 @@
 #define HnswGetMl(m) (1 / log(m))
 
 /* Ensure fits on page and in uint8 */
-#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / m) - 2, 255)
+#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
 {
-	List	   *heaptids;
+	slist_node	next;
+	ItemPointerData heaptids[HNSW_HEAPTIDS];
+	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
-	HnswNeighborArray *neighbors;
+	uint32		hash;
+	struct HnswNeighborArray *neighbors;
 	BlockNumber blkno;
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Vector	   *vec;
+	Datum		value;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -136,6 +148,59 @@ 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 */
@@ -159,16 +224,19 @@ typedef struct HnswBuildState
 	Oid			collation;
 
 	/* Variables */
-	List	   *elements;
+	HnswGraph	graphData;
-	HnswElement entryPoint;
+	HnswGraph  *graph;
 	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
@@ -204,7 +272,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		vec;
+	Vector		data;
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -250,7 +318,7 @@ typedef struct HnswVacuumState
 	Oid			collation;
 
 	/* Variables */
-	HTAB	   *deleted;
+	struct tidhash_hash *deleted;
 	BufferAccessStrategy bas;
 	HnswNeighborTuple ntup;
 	HnswElementData highestPoint;
@@ -264,31 +332,28 @@ int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
 bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
-void		HnswCommitBuffer(Buffer buf, GenericXLogState *state);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
-void		HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
 void		HnswInit(void);
 List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
 HnswElement HnswGetEntryPoint(Relation index);
 void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
-void		HnswFreeElement(HnswElement element);
 HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
 void		HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
-HnswElement HnswFindDuplicate(HnswElement e);
 HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
-void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum);
+void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
 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		HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building);
-void		HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting);
+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);
@@ -306,4 +371,31 @@ 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/hnswbuild.c

@@ -2,12 +2,16 @@
 
 #include <math.h>
 
+#include "access/parallel.h"
+#include "access/xact.h"
 #include "catalog/index.h"
 #include "hnsw.h"
 #include "miscadmin.h"
 #include "lib/pairingheap.h"
 #include "nodes/pg_list.h"
 #include "storage/bufmgr.h"
+#include "tcop/tcopprot.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -35,6 +39,27 @@
 #define UpdateProgress(index, val) ((void)val)
 #endif
 
+#if PG_VERSION_NUM >= 140000
+#include "utils/backend_status.h"
+#include "utils/wait_event.h"
+#endif
+
+#if PG_VERSION_NUM >= 120000
+#include "access/table.h"
+#include "optimizer/optimizer.h"
+#else
+#include "access/heapam.h"
+#include "optimizer/planner.h"
+#include "pgstat.h"
+#endif
+
+#define PARALLEL_KEY_HNSW_SHARED		UINT64CONST(0xA000000000000001)
+#define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000002)
+
+#if PG_VERSION_NUM < 130000
+#define GENERATIONCHUNK_RAWSIZE (SIZEOF_SIZE_T + SIZEOF_VOID_P * 2)
+#endif
+
 /*
  * Create the metapage
  */
@@ -45,11 +70,11 @@ CreateMetaPage(HnswBuildState * buildstate)
 	ForkNumber	forkNum = buildstate->forkNum;
 	Buffer		buf;
 	Page		page;
-	GenericXLogState *state;
 	HnswMetaPage metap;
 
 	buf = HnswNewBuffer(index, forkNum);
-	HnswInitRegisterPage(index, &buf, &page, &state);
+	page = BufferGetPage(buf);
+	HnswInitPage(buf, page);
 
 	/* Set metapage data */
 	metap = HnswPageGetMeta(page);
@@ -65,14 +90,15 @@ CreateMetaPage(HnswBuildState * buildstate)
 	((PageHeader) page)->pd_lower =
 		((char *) metap + sizeof(HnswMetaPageData)) - (char *) page;
 
-	HnswCommitBuffer(buf, state);
+	MarkBufferDirty(buf);
+	UnlockReleaseBuffer(buf);
 }
 
 /*
  * Add a new page
  */
 static void
-HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum)
+HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, ForkNumber forkNum)
 {
 	/* Add a new page */
 	Buffer		newbuf = HnswNewBuffer(index, forkNum);
@@ -81,7 +107,7 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
 	HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
 
 	/* Commit */
-	GenericXLogFinish(*state);
+	MarkBufferDirty(*buf);
 	UnlockReleaseBuffer(*buf);
 
 	/* Can take a while, so ensure we can interrupt */
@@ -92,8 +118,7 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
 
 	/* Prepare new page */
 	*buf = newbuf;
-	*state = GenericXLogStart(index);
+	*page = BufferGetPage(*buf);
-	*page = GenericXLogRegisterBuffer(*state, *buf, GENERIC_XLOG_FULL_IMAGE);
 	HnswInitPage(*buf, *page);
 }
 
@@ -105,46 +130,52 @@ CreateElementPages(HnswBuildState * buildstate)
 {
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
-	int			dimensions = buildstate->dimensions;
+	Size		etupAllocSize;
-	Size		etupSize;
 	Size		maxSize;
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
 	BlockNumber insertPage;
 	Buffer		buf;
 	Page		page;
-	GenericXLogState *state;
+	slist_iter	iter;
-	ListCell   *lc;
 
 	/* Calculate sizes */
+	etupAllocSize = BLCKSZ;
 	maxSize = HNSW_MAX_SIZE;
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
-	etup = palloc0(etupSize);
+	etup = palloc0(etupAllocSize);
 	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
 	buf = HnswNewBuffer(index, forkNum);
-	state = GenericXLogStart(index);
+	page = BufferGetPage(buf);
-	page = GenericXLogRegisterBuffer(state, buf, GENERIC_XLOG_FULL_IMAGE);
 	HnswInitPage(buf, page);
 
-	foreach(lc, buildstate->elements)
+	slist_foreach(iter, &buildstate->graph->elements)
 	{
-		HnswElement element = lfirst(lc);
+		HnswElement element = slist_container(HnswElementData, next, iter.cur);
+		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
 
-		HnswSetElementTuple(etup, element);
+		/* Zero memory for each element */
+		MemSet(etup, 0, etupAllocSize);
 
 		/* Calculate sizes */
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(element->value)));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
+		/* Initial size check */
+		if (etupSize > etupAllocSize)
+			elog(ERROR, "index tuple too large");
+
+		HnswSetElementTuple(etup, element);
+
 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
-			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
+			HnswBuildAppendPage(index, &buf, &page, forkNum);
 
 		/* Calculate offsets */
 		element->blkno = BufferGetBlockNumber(buf);
@@ -168,7 +199,7 @@ CreateElementPages(HnswBuildState * buildstate)
 
 		/* Add new page if needed */
 		if (PageGetFreeSpace(page) < ntupSize)
-			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
+			HnswBuildAppendPage(index, &buf, &page, forkNum);
 
 		/* Add placeholder for neighbors */
 		if (PageAddItem(page, (Item) ntup, ntupSize, InvalidOffsetNumber, false, false) != element->neighborOffno)
@@ -178,10 +209,10 @@ CreateElementPages(HnswBuildState * buildstate)
 	insertPage = BufferGetBlockNumber(buf);
 
 	/* Commit */
-	GenericXLogFinish(state);
+	MarkBufferDirty(buf);
 	UnlockReleaseBuffer(buf);
 
-	HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, buildstate->entryPoint, insertPage, forkNum);
+	HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, buildstate->graph->entryPoint, insertPage, forkNum, true);
 
 	pfree(etup);
 	pfree(ntup);
@@ -196,18 +227,17 @@ CreateNeighborPages(HnswBuildState * buildstate)
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
 	int			m = buildstate->m;
-	ListCell   *lc;
+	slist_iter	iter;
 	HnswNeighborTuple ntup;
 
 	/* Allocate once */
 	ntup = palloc0(BLCKSZ);
 
-	foreach(lc, buildstate->elements)
+	slist_foreach(iter, &buildstate->graph->elements)
 	{
-		HnswElement e = lfirst(lc);
+		HnswElement e = slist_container(HnswElementData, next, iter.cur);
 		Buffer		buf;
 		Page		page;
-		GenericXLogState *state;
 		Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 
 		/* Can take a while, so ensure we can interrupt */
@@ -216,8 +246,7 @@ CreateNeighborPages(HnswBuildState * buildstate)
 
 		buf = ReadBufferExtended(index, forkNum, e->neighborPage, RBM_NORMAL, NULL);
 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-		state = GenericXLogStart(index);
+		page = BufferGetPage(buf);
-		page = GenericXLogRegisterBuffer(state, buf, 0);
 
 		HnswSetNeighborTuple(ntup, e, m);
 
@@ -225,26 +254,30 @@ CreateNeighborPages(HnswBuildState * buildstate)
 			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 		/* Commit */
-		GenericXLogFinish(state);
+		MarkBufferDirty(buf);
 		UnlockReleaseBuffer(buf);
 	}
 
 	pfree(ntup);
 }
 
+#ifdef HNSW_MEMORY
 /*
- * Free elements
+ * Show memory usage
  */
 static void
-FreeElements(HnswBuildState * buildstate)
+ShowMemoryUsage(HnswBuildState * buildstate)
 {
-	ListCell   *lc;
+#if PG_VERSION_NUM >= 130000
-
+	elog(INFO, "graph memory: %zu MB, total memory: %zu MB",
-	foreach(lc, buildstate->elements)
+		 MemoryContextMemAllocated(buildstate->graphCtx, false) / (1024 * 1024),
-		HnswFreeElement(lfirst(lc));
+		 MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
-
+#else
-	list_free(buildstate->elements);
+	MemoryContextStats(CurrentMemoryContext);
+	elog(INFO, "estimated memory: %zu MB", buildstate->memoryUsed / (1024 * 1024));
+#endif
 }
+#endif
 
 /*
  * Flush pages
@@ -252,25 +285,78 @@ FreeElements(HnswBuildState * buildstate)
 static void
 FlushPages(HnswBuildState * buildstate)
 {
+#ifdef HNSW_MEMORY
+	ShowMemoryUsage(buildstate);
+#endif
+
 	CreateMetaPage(buildstate);
 	CreateElementPages(buildstate);
 	CreateNeighborPages(buildstate);
 
-	buildstate->flushed = true;
+	buildstate->graph->flushed = true;
-	FreeElements(buildstate);
+	MemoryContextReset(buildstate->graphCtx);
+}
+
+#if PG_VERSION_NUM < 130000
+/*
+ * Get the memory used by an element
+ */
+static long
+HnswElementMemory(HnswElement e, int m)
+{
+	long		elementSize = sizeof(HnswElementData);
+
+	elementSize += sizeof(HnswNeighborArray) * (e->level + 1);
+	elementSize += sizeof(HnswCandidate) * (m * (e->level + 2));
+	elementSize += VARSIZE_ANY(DatumGetPointer(e->value));
+	/* Each allocation has a chunk header */
+	elementSize += (e->level + 4) * GENERATIONCHUNK_RAWSIZE;
+	/* Add an extra 5% for alignment and other overhead */
+	return elementSize * 1.05;
+}
+#endif
+
+/*
+ * Find duplicate element
+ */
+static bool
+HnswFindDuplicateInMemory(HnswElement element)
+{
+	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;
+
+		/* Check for space */
+		if (neighbor->element->heaptidsLength < HNSW_HEAPTIDS)
+		{
+			HnswAddHeapTid(neighbor->element, &element->heaptids[0]);
+			return true;
+		}
+	}
+
+	return false;
 }
 
 /*
- * Insert tuple
+ * Insert tuple into in-memory graph
  */
 static bool
-InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup)
+InsertTupleInMemory(Relation index, Datum *values, ItemPointer heaptid, HnswBuildState * buildstate)
 {
 	FmgrInfo   *procinfo = buildstate->procinfo;
 	Oid			collation = buildstate->collation;
-	HnswElement entryPoint = buildstate->entryPoint;
+	HnswGraph  *graph = buildstate->graph;
+	HnswElement entryPoint = graph->entryPoint;
 	int			efConstruction = buildstate->efConstruction;
 	int			m = buildstate->m;
+	MemoryContext oldCtx;
+	HnswElement element;
 
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -282,35 +368,67 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 			return false;
 	}
 
-	/* Copy value to element so accessible outside of memory context */
+	/* Allocate element in graph memory context */
-	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
+	oldCtx = MemoryContextSwitchTo(buildstate->graphCtx);
+	element = HnswInitElement(heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel);
+	element->value = datumCopy(value, false, -1);
+	MemoryContextSwitchTo(oldCtx);
+
+	/* Update memory usage */
+#if PG_VERSION_NUM >= 130000
+	graph->memoryUsed = MemoryContextMemAllocated(buildstate->graphCtx, false);
+#else
+	graph->memoryUsed += HnswElementMemory(element, buildstate->m);
+#endif
 
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
 
 	/* Look for duplicate */
-	*dup = HnswFindDuplicate(element);
+	if (HnswFindDuplicateInMemory(element))
-
-	/* Update neighbors if needed */
-	if (*dup == NULL)
 	{
-		for (int lc = element->level; lc >= 0; lc--)
+		/* No need to free element since memory unlikely to be reallocated */
-		{
+		return true;
-			int			lm = HnswGetLayerM(m, lc);
+	}
-			HnswNeighborArray *neighbors = &element->neighbors[lc];
 
-			for (int i = 0; i < neighbors->length; i++)
+	/* Add element */
-				HnswUpdateConnection(element, &neighbors->items[i], lm, lc, NULL, NULL, procinfo, collation);
+	slist_push_head(&graph->elements, &element->next);
-		}
+
+	/* Update neighbors */
+	for (int lc = element->level; lc >= 0; lc--)
+	{
+		int			lm = HnswGetLayerM(m, lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(element, lc);
+
+		for (int i = 0; i < neighbors->length; i++)
+			HnswUpdateConnection(element, &neighbors->items[i], lm, lc, NULL, NULL, procinfo, collation);
 	}
 
 	/* Update entry point if needed */
-	if (*dup == NULL && (entryPoint == NULL || element->level > entryPoint->level))
+	if (entryPoint == NULL || element->level > entryPoint->level)
-		buildstate->entryPoint = element;
+		graph->entryPoint = element;
 
-	UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+	return true;
+}
 
-	return *dup == NULL;
+/*
+ * Acquire a lock if needed
+ */
+static inline void
+HnswLockAcquire(HnswShared * hnswshared)
+{
+	if (hnswshared)
+		SpinLockAcquire(&hnswshared->mutex);
+}
+
+/*
+ * Release a lock if needed
+ */
+static inline void
+HnswLockRelease(HnswShared * hnswshared)
+{
+	if (hnswshared)
+		SpinLockRelease(&hnswshared->mutex);
 }
 
 /*
@@ -321,9 +439,9 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 			  bool *isnull, bool tupleIsAlive, void *state)
 {
 	HnswBuildState *buildstate = (HnswBuildState *) state;
+	HnswGraph  *graph = buildstate->graph;
+	HnswShared *hnswshared = buildstate->hnswshared;
 	MemoryContext oldCtx;
-	HnswElement element;
-	HnswElement dup = NULL;
 	bool		inserted;
 
 #if PG_VERSION_NUM < 130000
@@ -334,69 +452,50 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (isnull[0])
 		return;
 
-	if (buildstate->indtuples >= buildstate->maxInMemoryElements)
+	/* Flush pages if needed */
+	if (!graph->flushed && graph->memoryUsed >= graph->memoryTotal)
 	{
-		if (!buildstate->flushed)
+		ereport(NOTICE,
-		{
+				(errmsg("hnsw graph no longer fits into maintenance_work_mem after " INT64_FORMAT " tuples", (int64) graph->indtuples),
-			ereport(NOTICE,
+				 errdetail("Building will take significantly more time."),
-					(errmsg("hnsw graph no longer fits into maintenance_work_mem after " INT64_FORMAT " tuples", (int64) buildstate->indtuples),
+				 errhint("Increase maintenance_work_mem to speed up builds.")));
-					 errdetail("Building will take significantly more time."),
-					 errhint("Increase maintenance_work_mem to speed up builds.")));
 
-			FlushPages(buildstate);
+		FlushPages(buildstate);
-		}
-
-		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
-
-		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
-			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
-
-		/* Reset memory context */
-		MemoryContextSwitchTo(oldCtx);
-		MemoryContextReset(buildstate->tmpCtx);
-
-		return;
 	}
 
-	/* Allocate necessary memory outside of memory context */
-	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
-	element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
-
-	/* Use memory context since detoast can allocate */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Insert tuple */
-	inserted = InsertTuple(index, values, element, buildstate, &dup);
+	if (graph->flushed)
+		inserted = HnswInsertTuple(index, values, isnull, tid, buildstate->heap, true);
+	else
+		inserted = InsertTupleInMemory(index, values, tid, buildstate);
+
+	/* Update progress */
+	if (inserted)
+	{
+		HnswLockAcquire(hnswshared);
+		UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++graph->indtuples);
+		HnswLockRelease(hnswshared);
+	}
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
 	MemoryContextReset(buildstate->tmpCtx);
-
-	/* Add outside memory context */
-	if (dup != NULL)
-		HnswAddHeapTid(dup, tid);
-
-	/* Add to buildstate or free */
-	if (inserted)
-		buildstate->elements = lappend(buildstate->elements, element);
-	else
-		HnswFreeElement(element);
 }
 
 /*
- * Get the max number of elements that fit into maintenance_work_mem
+ * Initialize the graph
  */
-static double
+static void
-HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
+InitGraph(HnswGraph * graph)
 {
-	Size		elementSize = sizeof(HnswElementData);
+	slist_init(&graph->elements);
-	double		avgLevel = -log(0.5) * ml;
+	graph->entryPoint = NULL;
-
+	graph->memoryUsed = 0;
-	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
+	graph->memoryTotal = maintenance_work_mem * 1024L;
-	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
+	graph->flushed = false;
-	elementSize += sizeof(ItemPointerData);
+	graph->indtuples = 0;
-	elementSize += VECTOR_SIZE(dimensions);
-	return (maintenance_work_mem * 1024L) / elementSize;
 }
 
 /*
@@ -432,19 +531,26 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
-	buildstate->elements = NIL;
+	InitGraph(&buildstate->graphData);
-	buildstate->entryPoint = NULL;
+	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->maxInMemoryElements = HnswGetMaxInMemoryElements(buildstate->m, buildstate->ml, buildstate->dimensions);
-	buildstate->flushed = false;
 
 	/* Reuse for each tuple */
 	buildstate->normvec = InitVector(buildstate->dimensions);
 
+	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
+												   "Hnsw build graph context",
+#if PG_VERSION_NUM >= 150000
+												   1024 * 1024, 1024 * 1024,
+#endif
+												   1024 * 1024);
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
+
+	buildstate->hnswleader = NULL;
+	buildstate->hnswshared = NULL;
 }
 
 /*
@@ -454,26 +560,427 @@ static void
 FreeBuildState(HnswBuildState * buildstate)
 {
 	pfree(buildstate->normvec);
+	MemoryContextDelete(buildstate->graphCtx);
 	MemoryContextDelete(buildstate->tmpCtx);
 }
 
+/*
+ * Within leader, wait for end of heap scan
+ */
+static double
+ParallelHeapScan(HnswBuildState * buildstate)
+{
+	HnswShared *hnswshared = buildstate->hnswleader->hnswshared;
+	int			nparticipanttuplesorts;
+	double		reltuples;
+
+	nparticipanttuplesorts = buildstate->hnswleader->nparticipanttuplesorts;
+	for (;;)
+	{
+		SpinLockAcquire(&hnswshared->mutex);
+		if (hnswshared->nparticipantsdone == nparticipanttuplesorts)
+		{
+			buildstate->graph = &hnswshared->graphData;
+			reltuples = hnswshared->reltuples;
+			SpinLockRelease(&hnswshared->mutex);
+			break;
+		}
+		SpinLockRelease(&hnswshared->mutex);
+
+		ConditionVariableSleep(&hnswshared->workersdonecv,
+							   WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
+	}
+
+	ConditionVariableCancelSleep();
+
+	return reltuples;
+}
+
+/*
+ * Perform a worker's portion of a parallel insert
+ */
+static void
+HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool progress)
+{
+	HnswBuildState buildstate;
+#if PG_VERSION_NUM >= 120000
+	TableScanDesc scan;
+#else
+	HeapScanDesc scan;
+#endif
+	double		reltuples;
+	IndexInfo  *indexInfo;
+
+	/* Join parallel scan */
+	indexInfo = BuildIndexInfo(hnswspool->index);
+	indexInfo->ii_Concurrent = hnswshared->isconcurrent;
+	InitBuildState(&buildstate, hnswspool->heap, hnswspool->index, indexInfo, MAIN_FORKNUM);
+	buildstate.graph = &hnswshared->graphData;
+	buildstate.hnswshared = hnswshared;
+#if PG_VERSION_NUM >= 120000
+	scan = table_beginscan_parallel(hnswspool->heap,
+									ParallelTableScanFromHnswShared(hnswshared));
+	reltuples = table_index_build_scan(hnswspool->heap, hnswspool->index, indexInfo,
+									   true, progress, BuildCallback,
+									   (void *) &buildstate, scan);
+#else
+	scan = heap_beginscan_parallel(hnswspool->heap, &hnswshared->heapdesc);
+	reltuples = IndexBuildHeapScan(hnswspool->heap, hnswspool->index, indexInfo,
+								   true, BuildCallback,
+								   (void *) &buildstate, scan);
+#endif
+
+	/* Record statistics */
+	SpinLockAcquire(&hnswshared->mutex);
+	hnswshared->nparticipantsdone++;
+	hnswshared->reltuples += reltuples;
+	SpinLockRelease(&hnswshared->mutex);
+
+	/* Log statistics */
+	if (progress)
+		ereport(DEBUG1, (errmsg("leader processed " INT64_FORMAT " tuples", (int64) reltuples)));
+	else
+		ereport(DEBUG1, (errmsg("worker processed " INT64_FORMAT " tuples", (int64) reltuples)));
+
+	/* Notify leader */
+	ConditionVariableSignal(&hnswshared->workersdonecv);
+
+	FreeBuildState(&buildstate);
+}
+
+/*
+ * Perform work within a launched parallel process
+ */
+void
+HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
+{
+	char	   *sharedquery;
+	HnswSpool  *hnswspool;
+	HnswShared *hnswshared;
+	Relation	heapRel;
+	Relation	indexRel;
+	LOCKMODE	heapLockmode;
+	LOCKMODE	indexLockmode;
+
+	/* Set debug_query_string for individual workers first */
+	sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
+	debug_query_string = sharedquery;
+
+	/* Report the query string from leader */
+	pgstat_report_activity(STATE_RUNNING, debug_query_string);
+
+	/* Look up shared state */
+	hnswshared = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_SHARED, false);
+
+	/* Open relations using lock modes known to be obtained by index.c */
+	if (!hnswshared->isconcurrent)
+	{
+		heapLockmode = ShareLock;
+		indexLockmode = AccessExclusiveLock;
+	}
+	else
+	{
+		heapLockmode = ShareUpdateExclusiveLock;
+		indexLockmode = RowExclusiveLock;
+	}
+
+	/* Open relations within worker */
+#if PG_VERSION_NUM >= 120000
+	heapRel = table_open(hnswshared->heaprelid, heapLockmode);
+#else
+	heapRel = heap_open(hnswshared->heaprelid, heapLockmode);
+#endif
+	indexRel = index_open(hnswshared->indexrelid, indexLockmode);
+
+	/* Initialize worker's own spool */
+	hnswspool = (HnswSpool *) palloc0(sizeof(HnswSpool));
+	hnswspool->heap = heapRel;
+	hnswspool->index = indexRel;
+
+	/* Perform inserts */
+	HnswParallelScanAndInsert(hnswspool, hnswshared, false);
+
+	/* Close relations within worker */
+	index_close(indexRel, indexLockmode);
+#if PG_VERSION_NUM >= 120000
+	table_close(heapRel, heapLockmode);
+#else
+	heap_close(heapRel, heapLockmode);
+#endif
+}
+
+/*
+ * End parallel build
+ */
+static void
+HnswEndParallel(HnswLeader * hnswleader)
+{
+	/* Shutdown worker processes */
+	WaitForParallelWorkersToFinish(hnswleader->pcxt);
+
+	/* Free last reference to MVCC snapshot, if one was used */
+	if (IsMVCCSnapshot(hnswleader->snapshot))
+		UnregisterSnapshot(hnswleader->snapshot);
+	DestroyParallelContext(hnswleader->pcxt);
+	ExitParallelMode();
+}
+
+/*
+ * Return size of shared memory required for parallel index build
+ */
+static Size
+ParallelEstimateShared(Relation heap, Snapshot snapshot)
+{
+#if PG_VERSION_NUM >= 120000
+	return add_size(BUFFERALIGN(sizeof(HnswShared)), table_parallelscan_estimate(heap, snapshot));
+#else
+	if (!IsMVCCSnapshot(snapshot))
+	{
+		Assert(snapshot == SnapshotAny);
+		return sizeof(HnswShared);
+	}
+
+	return add_size(offsetof(HnswShared, heapdesc) +
+					offsetof(ParallelHeapScanDescData, phs_snapshot_data),
+					EstimateSnapshotSpace(snapshot));
+#endif
+}
+
+/*
+ * Within leader, participate as a parallel worker
+ */
+static void
+HnswLeaderParticipateAsWorker(HnswBuildState * buildstate)
+{
+	HnswLeader *hnswleader = buildstate->hnswleader;
+	HnswSpool  *leaderworker;
+
+	/* Allocate memory and initialize private spool */
+	leaderworker = (HnswSpool *) palloc0(sizeof(HnswSpool));
+	leaderworker->heap = buildstate->heap;
+	leaderworker->index = buildstate->index;
+
+	/* Perform work common to all participants */
+	HnswParallelScanAndInsert(leaderworker, hnswleader->hnswshared, true);
+}
+
+/*
+ * Begin parallel build
+ */
+static void
+HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
+{
+	ParallelContext *pcxt;
+	int			scantuplesortstates;
+	Snapshot	snapshot;
+	Size		esthnswshared;
+	HnswShared *hnswshared;
+	HnswLeader *hnswleader = (HnswLeader *) palloc0(sizeof(HnswLeader));
+	bool		leaderparticipates = true;
+	int			querylen;
+
+#ifdef DISABLE_LEADER_PARTICIPATION
+	leaderparticipates = false;
+#endif
+
+	/* Enter parallel mode and create context */
+	EnterParallelMode();
+	Assert(request > 0);
+#if PG_VERSION_NUM >= 120000
+	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request);
+#else
+	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request, true);
+#endif
+
+	scantuplesortstates = leaderparticipates ? request + 1 : request;
+
+	/* Get snapshot for table scan */
+	if (!isconcurrent)
+		snapshot = SnapshotAny;
+	else
+		snapshot = RegisterSnapshot(GetTransactionSnapshot());
+
+	/* Estimate size of workspaces */
+	esthnswshared = ParallelEstimateShared(buildstate->heap, snapshot);
+	shm_toc_estimate_chunk(&pcxt->estimator, esthnswshared);
+	shm_toc_estimate_keys(&pcxt->estimator, 1);
+
+	/* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
+	if (debug_query_string)
+	{
+		querylen = strlen(debug_query_string);
+		shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
+		shm_toc_estimate_keys(&pcxt->estimator, 1);
+	}
+	else
+		querylen = 0;			/* keep compiler quiet */
+
+	/* Everyone's had a chance to ask for space, so now create the DSM */
+	InitializeParallelDSM(pcxt);
+
+	/* If no DSM segment was available, back out (do serial build) */
+	if (pcxt->seg == NULL)
+	{
+		if (IsMVCCSnapshot(snapshot))
+			UnregisterSnapshot(snapshot);
+		DestroyParallelContext(pcxt);
+		ExitParallelMode();
+		return;
+	}
+
+	/* Store shared build state, for which we reserved space */
+	hnswshared = (HnswShared *) shm_toc_allocate(pcxt->toc, esthnswshared);
+	/* Initialize immutable state */
+	hnswshared->heaprelid = RelationGetRelid(buildstate->heap);
+	hnswshared->indexrelid = RelationGetRelid(buildstate->index);
+	hnswshared->isconcurrent = isconcurrent;
+	hnswshared->scantuplesortstates = scantuplesortstates;
+	ConditionVariableInit(&hnswshared->workersdonecv);
+	SpinLockInit(&hnswshared->mutex);
+	/* Initialize mutable state */
+	hnswshared->nparticipantsdone = 0;
+	hnswshared->reltuples = 0;
+	InitGraph(&hnswshared->graphData);
+	/* TODO Support in-memory builds */
+	hnswshared->graphData.memoryTotal = 0;
+	hnswshared->graphData.flushed = true;
+#if PG_VERSION_NUM >= 120000
+	table_parallelscan_initialize(buildstate->heap,
+								  ParallelTableScanFromHnswShared(hnswshared),
+								  snapshot);
+#else
+	heap_parallelscan_initialize(&hnswshared->heapdesc, buildstate->heap, snapshot);
+#endif
+
+	shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
+
+	/* Store query string for workers */
+	if (debug_query_string)
+	{
+		char	   *sharedquery;
+
+		sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
+		memcpy(sharedquery, debug_query_string, querylen + 1);
+		shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
+	}
+
+	/* Launch workers, saving status for leader/caller */
+	LaunchParallelWorkers(pcxt);
+	hnswleader->pcxt = pcxt;
+	hnswleader->nparticipanttuplesorts = pcxt->nworkers_launched;
+	if (leaderparticipates)
+		hnswleader->nparticipanttuplesorts++;
+	hnswleader->hnswshared = hnswshared;
+	hnswleader->snapshot = snapshot;
+
+	/* If no workers were successfully launched, back out (do serial build) */
+	if (pcxt->nworkers_launched == 0)
+	{
+		HnswEndParallel(hnswleader);
+		return;
+	}
+
+	/* Log participants */
+	ereport(DEBUG1, (errmsg("using %d parallel workers", pcxt->nworkers_launched)));
+
+	/* Save leader state now that it's clear build will be parallel */
+	buildstate->hnswleader = hnswleader;
+
+	/* Join heap scan ourselves */
+	if (leaderparticipates)
+		HnswLeaderParticipateAsWorker(buildstate);
+
+	/* Wait for all launched workers */
+	WaitForParallelWorkersToAttach(pcxt);
+}
+
+/*
+ * Compute parallel workers
+ */
+static int
+ComputeParallelWorkers(Relation heap, Relation index)
+{
+	int			parallel_workers;
+
+	/* Make sure it's safe to use parallel workers */
+	parallel_workers = plan_create_index_workers(RelationGetRelid(heap), RelationGetRelid(index));
+	if (parallel_workers == 0)
+		return 0;
+
+	/* Use parallel_workers storage parameter on table if set */
+	parallel_workers = RelationGetParallelWorkers(heap, -1);
+	if (parallel_workers != -1)
+		return Min(parallel_workers, max_parallel_maintenance_workers);
+
+	return max_parallel_maintenance_workers;
+}
+
 /*
  * Build graph
  */
 static void
 BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
 {
+	int			parallel_workers = 0;
+
 	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_HNSW_PHASE_LOAD);
 
+	/* Calculate parallel workers */
+	if (buildstate->heap != NULL && hnsw_enable_parallel_build)
+		parallel_workers = ComputeParallelWorkers(buildstate->heap, buildstate->index);
+
+	/* Attempt to launch parallel worker scan when required */
+	if (parallel_workers > 0)
+	{
+		/* TODO Support in-memory builds */
+		FlushPages(buildstate);
+		HnswBeginParallel(buildstate, buildstate->indexInfo->ii_Concurrent, parallel_workers);
+	}
+
+	/* Add tuples to graph */
+	if (buildstate->heap != NULL)
+	{
+		if (buildstate->hnswleader)
+			buildstate->reltuples = ParallelHeapScan(buildstate);
+		else
+		{
 #if PG_VERSION_NUM >= 120000
-	buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+			buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, true, BuildCallback, (void *) buildstate, NULL);
+														   true, true, BuildCallback, (void *) buildstate, NULL);
 #else
-	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+			buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-											   true, BuildCallback, (void *) buildstate, NULL);
+													   true, BuildCallback, (void *) buildstate, NULL);
 #endif
+		}
+
+		buildstate->indtuples = buildstate->graph->indtuples;
+	}
+
+	/* Flush pages */
+	if (!buildstate->graph->flushed)
+		FlushPages(buildstate);
+
+	/* End parallel build */
+	if (buildstate->hnswleader)
+		HnswEndParallel(buildstate->hnswleader);
 }
 
+#if PG_VERSION_NUM < 110008
+void
+log_newpage_range(Relation rel, ForkNumber forkNum, BlockNumber startblk, BlockNumber endblk, bool page_std)
+{
+	for (BlockNumber blkno = startblk; blkno < endblk; blkno++)
+	{
+		Buffer		buf = ReadBufferExtended(rel, forkNum, blkno, RBM_NORMAL, NULL);
+
+		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
+		MarkBufferDirty(buf);
+		log_newpage_buffer(buf, page_std);
+		UnlockReleaseBuffer(buf);
+	}
+}
+#endif
+
 /*
  * Build the index
  */
@@ -481,13 +988,16 @@ static void
 BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
 		   HnswBuildState * buildstate, ForkNumber forkNum)
 {
+#ifdef HNSW_MEMORY
+	SeedRandom(42);
+#endif
+
 	InitBuildState(buildstate, heap, index, indexInfo, forkNum);
 
-	if (buildstate->heap != NULL)
+	BuildGraph(buildstate, forkNum);
-		BuildGraph(buildstate, forkNum);
 
-	if (!buildstate->flushed)
+	if (RelationNeedsWAL(index))
-		FlushPages(buildstate);
+		log_newpage_range(index, forkNum, 0, RelationGetNumberOfBlocks(index), true);
 
 	FreeBuildState(buildstate);
 }

src/hnswinsert.c

@@ -5,6 +5,7 @@
 #include "hnsw.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 
 /*
@@ -92,7 +93,7 @@ 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)
+HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page, bool building)
 {
 	/* Add a new page */
 	LockRelationForExtension(index, ExclusiveLock);
@@ -100,7 +101,11 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
 	UnlockRelationForExtension(index, ExclusiveLock);
 
 	/* Init new page */
-	*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
+	if (building)
+		*npage = BufferGetPage(*nbuf);
+	else
+		*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
+
 	HnswInitPage(*nbuf, *npage);
 
 	/* Update previous buffer */
@@ -111,7 +116,7 @@ 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)
+WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage, bool building)
 {
 	Buffer		buf;
 	Page		page;
@@ -123,7 +128,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	Size		minCombinedSize;
 	HnswElementTuple etup;
 	BlockNumber currentPage = insertPage;
-	int			dimensions = e->vec->dim;
 	HnswNeighborTuple ntup;
 	Buffer		nbuf;
 	Page		npage;
@@ -132,7 +136,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -152,8 +156,16 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 		buf = ReadBuffer(index, currentPage);
 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
 
-		state = GenericXLogStart(index);
+		if (building)
-		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 */
 		if (!BlockNumberIsValid(newInsertPage) && PageGetFreeSpace(page) >= minCombinedSize)
@@ -173,7 +185,12 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 		if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
 		{
 			if (nbuf != buf)
-				npage = GenericXLogRegisterBuffer(state, nbuf, 0);
+			{
+				if (building)
+					npage = BufferGetPage(nbuf);
+				else
+					npage = GenericXLogRegisterBuffer(state, nbuf, 0);
+			}
 
 			break;
 		}
@@ -182,7 +199,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 		/* 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);
+			HnswInsertAppendPage(index, &nbuf, &npage, state, page, building);
 			break;
 		}
 
@@ -191,7 +208,8 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 		if (BlockNumberIsValid(currentPage))
 		{
 			/* Move to next page */
-			GenericXLogAbort(state);
+			if (!building)
+				GenericXLogAbort(state);
 			UnlockReleaseBuffer(buf);
 		}
 		else
@@ -199,22 +217,33 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 			Buffer		newbuf;
 			Page		newpage;
 
-			HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
+			HnswInsertAppendPage(index, &newbuf, &newpage, state, page, building);
 
 			/* Commit */
-			GenericXLogFinish(state);
+			if (building)
+				MarkBufferDirty(buf);
+			else
+				GenericXLogFinish(state);
 
 			/* Unlock previous buffer */
 			UnlockReleaseBuffer(buf);
 
 			/* Prepare new buffer */
-			state = GenericXLogStart(index);
 			buf = newbuf;
-			page = GenericXLogRegisterBuffer(state, buf, 0);
+			if (building)
+			{
+				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);
+				HnswInsertAppendPage(index, &nbuf, &npage, state, page, building);
 			else
 			{
 				nbuf = buf;
@@ -268,7 +297,14 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	}
 
 	/* Commit */
-	GenericXLogFinish(state);
+	if (building)
+	{
+		MarkBufferDirty(buf);
+		if (nbuf != buf)
+			MarkBufferDirty(nbuf);
+	}
+	else
+		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 	if (nbuf != buf)
 		UnlockReleaseBuffer(nbuf);
@@ -302,12 +338,12 @@ ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
  * Update neighbors
  */
 void
-HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting)
+HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
 {
 	for (int lc = e->level; lc >= 0; lc--)
 	{
 		int			lm = HnswGetLayerM(m, lc);
-		HnswNeighborArray *neighbors = &e->neighbors[lc];
+		HnswNeighborArray *neighbors = HnswGetNeighbors(e, lc);
 
 		for (int i = 0; i < neighbors->length; i++)
 		{
@@ -343,8 +379,16 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 			/* Register page */
 			buf = ReadBuffer(index, hc->element->neighborPage);
 			LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-			state = GenericXLogStart(index);
+			if (building)
-			page = GenericXLogRegisterBuffer(state, buf, 0);
+			{
+				state = NULL;
+				page = BufferGetPage(buf);
+			}
+			else
+			{
+				state = GenericXLogStart(index);
+				page = GenericXLogRegisterBuffer(state, buf, 0);
+			}
 
 			/* Get tuple */
 			itemid = PageGetItemId(page, offno);
@@ -386,9 +430,12 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 					elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 				/* Commit */
-				GenericXLogFinish(state);
+				if (building)
+					MarkBufferDirty(buf);
+				else
+					GenericXLogFinish(state);
 			}
-			else
+			else if (!building)
 				GenericXLogAbort(state);
 
 			UnlockReleaseBuffer(buf);
@@ -400,23 +447,34 @@ 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)
+HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool building)
 {
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
+	ItemId		itemid;
 	HnswElementTuple etup;
+	Size		etupSize;
 	int			i;
 
 	/* Read page */
 	buf = ReadBuffer(index, dup->blkno);
 	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-	state = GenericXLogStart(index);
+	if (building)
-	page = GenericXLogRegisterBuffer(state, buf, 0);
+	{
+		state = NULL;
+		page = BufferGetPage(buf);
+	}
+	else
+	{
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+	}
 
 	/* Find space */
-	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
+	itemid = PageGetItemId(page, dup->offno);
+	etup = (HnswElementTuple) PageGetItem(page, itemid);
+	etupSize = ItemIdGetLength(itemid);
 	for (i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (!ItemPointerIsValid(&etup->heaptids[i]))
@@ -426,60 +484,84 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	/* Either being deleted or we lost our chance to another backend */
 	if (i == 0 || i == HNSW_HEAPTIDS)
 	{
-		GenericXLogAbort(state);
+		if (!building)
+			GenericXLogAbort(state);
 		UnlockReleaseBuffer(buf);
 		return false;
 	}
 
 	/* Add heap TID */
-	etup->heaptids[i] = *((ItemPointer) linitial(element->heaptids));
+	etup->heaptids[i] = element->heaptids[0];
 
 	/* Overwrite tuple */
 	if (!PageIndexTupleOverwrite(page, dup->offno, (Item) etup, etupSize))
 		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
 
 	/* Commit */
-	GenericXLogFinish(state);
+	if (building)
+		MarkBufferDirty(buf);
+	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 dup, HnswElement entryPoint)
+WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
-	/* Try to add to existing page */
+	/* Look for duplicate */
-	if (dup != NULL)
+	if (HnswFindDuplicate(index, element, building))
-	{
+		return;
-		if (HnswAddDuplicate(index, element, dup))
-			return;
-	}
 
 	/* Write element and neighbor tuples */
-	WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage);
+	WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage, building);
 
 	/* Update insert page if needed */
 	if (BlockNumberIsValid(newInsertPage))
-		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM);
+		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);
 
 	/* Update neighbors */
-	HnswUpdateNeighborPages(index, procinfo, collation, element, m, false);
+	HnswUpdateNeighborPages(index, procinfo, collation, element, m, false, building);
 
-	/* Update metapage if needed */
+	/* Update entry point if needed */
 	if (entryPoint == NULL || element->level > entryPoint->level)
-		HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
+		HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, building);
 }
 
 /*
  * Insert a tuple into the index
  */
 bool
-HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building)
 {
 	Datum		value;
 	FmgrInfo   *normprocinfo;
@@ -489,7 +571,6 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	int			efConstruction = HnswGetEfConstruction(index);
 	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	Oid			collation = index->rd_indcollation[0];
-	HnswElement dup;
 	LOCKMODE	lockmode = ShareLock;
 
 	/* Detoast once for all calls */
@@ -515,7 +596,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 
 	/* Create an element */
 	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->vec = DatumGetVector(value);
+	element->value = value;
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -534,11 +615,8 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, false);
 
-	/* Look for duplicate */
-	dup = HnswFindDuplicate(element);
-
 	/* Write to disk */
-	WriteElement(index, procinfo, collation, element, m, efConstruction, dup, entryPoint);
+	WriteElement(index, procinfo, collation, element, m, efConstruction, entryPoint, building);
 
 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -572,7 +650,7 @@ hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 	oldCtx = MemoryContextSwitchTo(insertCtx);
 
 	/* Insert tuple */
-	HnswInsertTuple(index, values, isnull, heap_tid, heap);
+	HnswInsertTuple(index, values, isnull, heap_tid, heap, false);
 
 	/* Delete memory context */
 	MemoryContextSwitchTo(oldCtx);

src/hnswscan.c

@@ -188,15 +188,13 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		ItemPointer heaptid;
 
 		/* Move to next element if no valid heap TIDs */
-		if (list_length(hc->element->heaptids) == 0)
+		if (hc->element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
 			continue;
 		}
 
-		heaptid = llast(hc->element->heaptids);
+		heaptid = &hc->element->heaptids[--hc->element->heaptidsLength];
-
-		hc->element->heaptids = list_delete_last(hc->element->heaptids);
 
 		MemoryContextSwitchTo(oldCtx);
 

src/hnswutils.c

@@ -4,8 +4,92 @@
 
 #include "hnsw.h"
 #include "storage/bufmgr.h"
+#include "utils/datum.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 130000
+#include "common/hashfn.h"
+#else
+#include "utils/hashutils.h"
+#endif
+
+#if PG_VERSION_NUM < 170000
+static inline uint64
+murmurhash64(uint64 data)
+{
+	uint64		h = data;
+
+	h ^= h >> 33;
+	h *= 0xff51afd7ed558ccd;
+	h ^= h >> 33;
+	h *= 0xc4ceb9fe1a85ec53;
+	h ^= h >> 33;
+
+	return h;
+}
+#endif
+
+/* TID hash table */
+static uint32
+hash_tid(ItemPointerData tid)
+{
+	union
+	{
+		uint64		i;
+		ItemPointerData tid;
+	}			x;
+
+	/* Initialize unused bytes */
+	x.i = 0;
+	x.tid = tid;
+
+	return murmurhash64(x.i);
+}
+
+#define SH_PREFIX		tidhash
+#define SH_ELEMENT_TYPE	TidHashEntry
+#define SH_KEY_TYPE		ItemPointerData
+#define	SH_KEY			tid
+#define SH_HASH_KEY(tb, key)	hash_tid(key)
+#define SH_EQUAL(tb, a, b)		ItemPointerEquals(&a, &b)
+#define	SH_SCOPE		extern
+#define SH_DEFINE
+#include "lib/simplehash.h"
+
+/* Needed to include simplehash.h twice */
+#if PG_VERSION_NUM < 120000
+#undef SH_EQUAL
+#define sh_log2 pointerhash_sh_log2
+#define sh_pow2 pointerhash_sh_pow2
+#endif
+
+/* Pointer hash table */
+static uint32
+hash_pointer(uintptr_t ptr)
+{
+#if SIZEOF_VOID_P == 8
+	return murmurhash64((uint64) ptr);
+#else
+	return murmurhash32((uint32) ptr);
+#endif
+}
+
+#define SH_PREFIX		pointerhash
+#define SH_ELEMENT_TYPE	PointerHashEntry
+#define SH_KEY_TYPE		uintptr_t
+#define	SH_KEY			ptr
+#define SH_HASH_KEY(tb, key)	hash_pointer(key)
+#define SH_EQUAL(tb, a, b)		(a == b)
+#define	SH_SCOPE		extern
+#define SH_DEFINE
+#include "lib/simplehash.h"
+
+typedef union
+{
+	pointerhash_hash *pointers;
+	tidhash_hash *tids;
+}			visited_hash;
+
 /*
  * Get the max number of connections in an upper layer for each element in the index
  */
@@ -100,27 +184,6 @@ HnswInitPage(Buffer buf, Page page)
 	HnswPageGetOpaque(page)->page_id = HNSW_PAGE_ID;
 }
 
-/*
- * Init and register page
- */
-void
-HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state)
-{
-	*state = GenericXLogStart(index);
-	*page = GenericXLogRegisterBuffer(*state, *buf, GENERIC_XLOG_FULL_IMAGE);
-	HnswInitPage(*buf, *page);
-}
-
-/*
- * Commit buffer
- */
-void
-HnswCommitBuffer(Buffer buf, GenericXLogState *state)
-{
-	GenericXLogFinish(state);
-	UnlockReleaseBuffer(buf);
-}
-
 /*
  * Allocate neighbors
  */
@@ -143,17 +206,6 @@ HnswInitNeighbors(HnswElement element, int m)
 	}
 }
 
-/*
- * Free neighbors
- */
-static void
-HnswFreeNeighbors(HnswElement element)
-{
-	for (int lc = 0; lc <= element->level; lc++)
-		pfree(element->neighbors[lc].items);
-	pfree(element->neighbors);
-}
-
 /*
  * Allocate an element
  */
@@ -168,7 +220,7 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
 	if (level > maxLevel)
 		level = maxLevel;
 
-	element->heaptids = NIL;
+	element->heaptidsLength = 0;
 	HnswAddHeapTid(element, heaptid);
 
 	element->level = level;
@@ -176,19 +228,9 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
 
 	HnswInitNeighbors(element, m);
 
-	return element;
+	element->value = PointerGetDatum(NULL);
-}
 
-/*
+	return element;
- * Free an element
- */
-void
-HnswFreeElement(HnswElement element)
-{
-	HnswFreeNeighbors(element);
-	list_free_deep(element->heaptids);
-	pfree(element->vec);
-	pfree(element);
 }
 
 /*
@@ -197,10 +239,7 @@ HnswFreeElement(HnswElement element)
 void
 HnswAddHeapTid(HnswElement element, ItemPointer heaptid)
 {
-	ItemPointer copy = palloc(sizeof(ItemPointerData));
+	element->heaptids[element->heaptidsLength++] = *heaptid;
-
-	ItemPointerCopy(heaptid, copy);
-	element->heaptids = lappend(element->heaptids, copy);
 }
 
 /*
@@ -214,7 +253,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->vec = NULL;
+	element->value = PointerGetDatum(NULL);
 	return element;
 }
 
@@ -292,7 +331,7 @@ HnswUpdateMetaPageInfo(Page page, int updateEntry, HnswElement entryPoint, Block
  * Update the metapage
  */
 void
-HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum)
+HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building)
 {
 	Buffer		buf;
 	Page		page;
@@ -300,12 +339,24 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 
 	buf = ReadBufferExtended(index, forkNum, HNSW_METAPAGE_BLKNO, RBM_NORMAL, NULL);
 	LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
-	state = GenericXLogStart(index);
+	if (building)
-	page = GenericXLogRegisterBuffer(state, buf, 0);
+	{
+		state = NULL;
+		page = BufferGetPage(buf);
+	}
+	else
+	{
+		state = GenericXLogStart(index);
+		page = GenericXLogRegisterBuffer(state, buf, 0);
+	}
 
 	HnswUpdateMetaPageInfo(page, updateEntry, entryPoint, insertPage);
 
-	HnswCommitBuffer(buf, state);
+	if (building)
+		MarkBufferDirty(buf);
+	else
+		GenericXLogFinish(state);
+	UnlockReleaseBuffer(buf);
 }
 
 /*
@@ -319,12 +370,12 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
 	etup->deleted = 0;
 	for (int i = 0; i < HNSW_HEAPTIDS; i++)
 	{
-		if (i < list_length(element->heaptids))
+		if (i < element->heaptidsLength)
-			etup->heaptids[i] = *((ItemPointer) list_nth(element->heaptids, i));
+			etup->heaptids[i] = element->heaptids[i];
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
+	memcpy(&etup->data, DatumGetPointer(element->value), VARSIZE_ANY(DatumGetPointer(element->value)));
 }
 
 /*
@@ -339,7 +390,7 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
 
 	for (int lc = e->level; lc >= 0; lc--)
 	{
-		HnswNeighborArray *neighbors = &e->neighbors[lc];
+		HnswNeighborArray *neighbors = HnswGetNeighbors(e, lc);
 		int			lm = HnswGetLayerM(m, lc);
 
 		for (int i = 0; i < lm; i++)
@@ -397,7 +448,7 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
 		if (level < 0)
 			level = 0;
 
-		neighbors = &element->neighbors[level];
+		neighbors = HnswGetNeighbors(element, level);
 		hc = &neighbors->items[neighbors->length++];
 		hc->element = e;
 	}
@@ -431,7 +482,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	element->deleted = etup->deleted;
 	element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
 	element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
-	element->heaptids = NIL;
+	element->heaptidsLength = 0;
 
 	if (loadHeaptids)
 	{
@@ -446,10 +497,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	}
 
 	if (loadVec)
-	{
+		element->value = datumCopy(PointerGetDatum(&etup->data), false, -1);
-		element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
-		memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
-	}
 }
 
 /*
@@ -476,7 +524,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
+		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
 
 	UnlockReleaseBuffer(buf);
 }
@@ -487,7 +535,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 static float
 GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
 {
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
 }
 
 /*
@@ -552,16 +600,22 @@ CreatePairingHeapNode(HnswCandidate * c)
  * Add to visited
  */
 static inline void
-AddToVisited(HTAB *v, HnswCandidate * hc, Relation index, bool *found)
+AddToVisited(visited_hash v, HnswCandidate * hc, Relation index, bool *found)
 {
 	if (index == NULL)
-		hash_search(v, &hc->element, HASH_ENTER, found);
+	{
+#if PG_VERSION_NUM >= 130000
+		pointerhash_insert_hash(v.pointers, (uintptr_t) hc->element, hc->element->hash, found);
+#else
+		pointerhash_insert(v.pointers, (uintptr_t) hc->element, found);
+#endif
+	}
 	else
 	{
 		ItemPointerData indextid;
 
 		ItemPointerSet(&indextid, hc->element->blkno, hc->element->offno);
-		hash_search(v, &indextid, HASH_ENTER, found);
+		tidhash_insert(v.tids, indextid, found);
 	}
 }
 
@@ -571,36 +625,26 @@ AddToVisited(HTAB *v, HnswCandidate * hc, Relation index, bool *found)
 List *
 HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
 {
-	ListCell   *lc2;
-
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
 	pairingheap *W = pairingheap_allocate(CompareFurthestCandidates, NULL);
 	int			wlen = 0;
-	HASHCTL		hash_ctl;
+	visited_hash v;
-	HTAB	   *v;
+	ListCell   *lc2;
 
 	/* Create hash table */
 	if (index == NULL)
-	{
+		v.pointers = pointerhash_create(CurrentMemoryContext, ef * m * 2, NULL);
-		hash_ctl.keysize = sizeof(HnswElement *);
-		hash_ctl.entrysize = sizeof(HnswElement *);
-	}
 	else
-	{
+		v.tids = tidhash_create(CurrentMemoryContext, ef * m * 2, NULL);
-		hash_ctl.keysize = sizeof(ItemPointerData);
-		hash_ctl.entrysize = sizeof(ItemPointerData);
-	}
-
-	hash_ctl.hcxt = CurrentMemoryContext;
-	v = hash_create("hnsw visited", 256, &hash_ctl, HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
 
 	/* Add entry points to v, C, and W */
 	foreach(lc2, ep)
 	{
 		HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
+		bool		found;
 
-		AddToVisited(v, hc, index, NULL);
+		AddToVisited(v, hc, index, &found);
 
 		pairingheap_add(C, &(CreatePairingHeapNode(hc)->ph_node));
 		pairingheap_add(W, &(CreatePairingHeapNode(hc)->ph_node));
@@ -610,7 +654,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 		 * would be ideal to do this for inserts as well, but this could
 		 * affect insert performance.
 		 */
-		if (skipElement == NULL || list_length(hc->element->heaptids) != 0)
+		if (skipElement == NULL || hc->element->heaptidsLength != 0)
 			wlen++;
 	}
 
@@ -627,7 +671,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 			HnswLoadNeighbors(c->element, index, m);
 
 		/* Get the neighborhood at layer lc */
-		neighborhood = &c->element->neighbors[lc];
+		neighborhood = HnswGetNeighbors(c->element, lc);
 
 		for (int i = 0; i < neighborhood->length; i++)
 		{
@@ -669,7 +713,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 					 * vacuuming. It would be ideal to do this for inserts as
 					 * well, but this could affect insert performance.
 					 */
-					if (skipElement == NULL || list_length(e->element->heaptids) != 0)
+					if (skipElement == NULL || e->element->heaptidsLength != 0)
 					{
 						wlen++;
 
@@ -730,27 +774,27 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
 	/* Look for cached distance */
 	if (a->neighbors != NULL)
 	{
-		Assert(a->level >= lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(a, lc);
 
-		for (int i = 0; i < a->neighbors[lc].length; i++)
+		for (int i = 0; i < neighbors->length; i++)
 		{
-			if (a->neighbors[lc].items[i].element == b)
+			if (neighbors->items[i].element == b)
-				return a->neighbors[lc].items[i].distance;
+				return neighbors->items[i].distance;
 		}
 	}
 
 	if (b->neighbors != NULL)
 	{
-		Assert(b->level >= lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(b, lc);
 
-		for (int i = 0; i < b->neighbors[lc].length; i++)
+		for (int i = 0; i < neighbors->length; i++)
 		{
-			if (b->neighbors[lc].items[i].element == a)
+			if (neighbors->items[i].element == a)
-				return b->neighbors[lc].items[i].distance;
+				return neighbors->items[i].distance;
 		}
 	}
 
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
 }
 
 /*
@@ -782,7 +826,8 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
 	pairingheap *wd;
-	bool		mustCalculate = !e2->neighbors[lc].closerSet;
+	HnswNeighborArray *neighbors = HnswGetNeighbors(e2, lc);
+	bool		mustCalculate = !neighbors->closerSet;
 	List	   *added = NIL;
 	bool		removedAny = false;
 
@@ -846,7 +891,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 	}
 
 	/* Cached value can only be used in future if sorted deterministically */
-	e2->neighbors[lc].closerSet = sortCandidates;
+	neighbors->closerSet = sortCandidates;
 
 	/* Keep pruned connections */
 	while (!pairingheap_is_empty(wd) && list_length(r) < m)
@@ -864,30 +909,6 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 	return r;
 }
 
-/*
- * Find duplicate element
- */
-HnswElement
-HnswFindDuplicate(HnswElement e)
-{
-	HnswNeighborArray *neighbors = &e->neighbors[0];
-
-	for (int i = 0; i < neighbors->length; i++)
-	{
-		HnswCandidate *neighbor = &neighbors->items[i];
-
-		/* Exit early since ordered by distance */
-		if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
-			break;
-
-		/* Check for space */
-		if (list_length(neighbor->element->heaptids) < HNSW_HEAPTIDS)
-			return neighbor->element;
-	}
-
-	return NULL;
-}
-
 /*
  * Add connections
  */
@@ -895,7 +916,7 @@ static void
 AddConnections(HnswElement element, List *neighbors, int m, int lc)
 {
 	ListCell   *lc2;
-	HnswNeighborArray *a = &element->neighbors[lc];
+	HnswNeighborArray *a = HnswGetNeighbors(element, lc);
 
 	foreach(lc2, neighbors)
 		a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
@@ -907,7 +928,7 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
 void
 HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation)
 {
-	HnswNeighborArray *currentNeighbors = &hc->element->neighbors[lc];
+	HnswNeighborArray *currentNeighbors = HnswGetNeighbors(hc->element, lc);
 
 	HnswCandidate hc2;
 
@@ -930,19 +951,19 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = PointerGetDatum(hc->element->vec);
+			Datum		q = hc->element->value;
 
 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (hc3->element->vec == NULL)
+				if (DatumGetPointer(hc3->element->value) == NULL)
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
 
 				/* Prune element if being deleted */
-				if (list_length(hc3->element->heaptids) == 0)
+				if (hc3->element->heaptidsLength == 0)
 				{
 					pruned = &currentNeighbors->items[i];
 					break;
@@ -1000,7 +1021,7 @@ RemoveElements(List *w, HnswElement skipElement)
 		if (skipElement != NULL && hc->element->blkno == skipElement->blkno && hc->element->offno == skipElement->offno)
 			continue;
 
-		if (list_length(hc->element->heaptids) != 0)
+		if (hc->element->heaptidsLength != 0)
 			w2 = lappend(w2, hc);
 	}
 
@@ -1017,9 +1038,15 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = PointerGetDatum(element->vec);
+	Datum		q = element->value;
 	HnswElement skipElement = existing ? element : NULL;
 
+#if PG_VERSION_NUM >= 130000
+	/* Precompute hash */
+	if (index == NULL)
+		element->hash = hash_pointer((uintptr_t) element);
+#endif
+
 	/* No neighbors if no entry point */
 	if (entryPoint == NULL)
 		return;

src/hnswvacuum.c

@@ -12,12 +12,9 @@
  * Check if deleted list contains an index TID
  */
 static bool
-DeletedContains(HTAB *deleted, ItemPointer indextid)
+DeletedContains(tidhash_hash * deleted, ItemPointer indextid)
 {
-	bool		found;
+	return tidhash_lookup(deleted, *indextid) != NULL;
-
-	hash_search(deleted, indextid, HASH_FIND, &found);
-	return found;
 }
 
 /*
@@ -62,7 +59,8 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		/* Iterate over nodes */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			int			idx = 0;
 			bool		itemUpdated = false;
 
@@ -93,7 +91,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 				if (itemUpdated)
 				{
-					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+					Size		etupSize = ItemIdGetLength(itemid);
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -109,11 +107,13 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 			if (!ItemPointerIsValid(&etup->heaptids[0]))
 			{
 				ItemPointerData ip;
+				bool		found;
 
 				/* Add to deleted list */
 				ItemPointerSet(&ip, blkno, offno);
 
-				(void) hash_search(vacuumstate->deleted, &ip, HASH_ENTER, NULL);
+				tidhash_insert(vacuumstate->deleted, ip, &found);
+				Assert(!found);
 			}
 			else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
 			{
@@ -206,7 +206,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 
 	/* Init fields */
 	HnswInitNeighbors(element, m);
-	element->heaptids = NIL;
+	element->heaptidsLength = 0;
 
 	/* Add element to graph, skipping itself */
 	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, true);
@@ -230,7 +230,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	UnlockReleaseBuffer(buf);
 
 	/* Update neighbors */
-	HnswUpdateNeighborPages(index, procinfo, collation, element, m, true);
+	HnswUpdateNeighborPages(index, procinfo, collation, element, m, true, false);
 }
 
 /*
@@ -286,7 +286,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 			 * 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);
+			HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, highestPoint, InvalidBlockNumber, MAIN_FORKNUM, false);
 		}
 		else
 		{
@@ -419,7 +419,7 @@ RepairGraph(HnswVacuumState * vacuumstate)
 			 * was replaced and highest point was outdated.
 			 */
 			if (entryPoint == NULL || element->level > entryPoint->level)
-				HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM);
+				HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, false);
 
 			/* Release lock */
 			UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -477,7 +477,8 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		/* Update element and neighbors together */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
@@ -505,7 +506,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				continue;
 
 			/* Calculate sizes */
-			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+			etupSize = ItemIdGetLength(itemid);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -528,7 +529,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
+			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)
@@ -563,7 +564,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	}
 
 	/* Update insert page last, after everything has been marked as deleted */
-	HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM);
+	HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM, false);
 }
 
 /*
@@ -573,7 +574,6 @@ 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));
@@ -595,10 +595,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
 
 	/* Create hash table */
-	hash_ctl.keysize = sizeof(ItemPointerData);
+	vacuumstate->deleted = tidhash_create(CurrentMemoryContext, 256, NULL);
-	hash_ctl.entrysize = sizeof(ItemPointerData);
-	hash_ctl.hcxt = CurrentMemoryContext;
-	vacuumstate->deleted = hash_create("hnswbulkdelete indextids", 256, &hash_ctl, HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
 }
 
 /*
@@ -607,7 +604,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 static void
 FreeVacuumState(HnswVacuumState * vacuumstate)
 {
-	hash_destroy(vacuumstate->deleted);
+	tidhash_destroy(vacuumstate->deleted);
 	FreeAccessStrategy(vacuumstate->bas);
 	pfree(vacuumstate->ntup);
 	MemoryContextDelete(vacuumstate->tmpCtx);

src/ivfbuild.c

@@ -11,7 +11,6 @@
 #include "miscadmin.h"
 #include "storage/bufmgr.h"
 #include "tcop/tcopprot.h"
-#include "utils/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -66,18 +65,11 @@
 static void
 AddSample(Datum *values, IvfflatBuildState * buildstate)
 {
-	MemoryContext oldCtx;
+	VectorArray samples = buildstate->samples;
-	Datum		value;
+	int			targsamples = samples->maxlen;
-	int			targsamples = buildstate->targsamples;
-
-	/* Use memory context since detoast can allocate */
-	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Detoast once for all calls */
-	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
-
-	/* Restore memory context */
-	MemoryContextSwitchTo(oldCtx);
 
 	/*
 	 * Normalize with KMEANS_NORM_PROC since spherical distance function
@@ -89,23 +81,18 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 			return;
 	}
 
-	/* Copy datum */
+	if (samples->length < targsamples)
-	value = datumCopy(value, false, -1);
+	{
-
+		VectorArraySet(samples, samples->length, DatumGetVector(value));
-	/* Reset memory context */
+		samples->length++;
-	MemoryContextReset(buildstate->tmpCtx);
+	}
-
-	if (list_length(buildstate->samples) < targsamples)
-		buildstate->samples = lappend(buildstate->samples, DatumGetVector(value));
 	else
 	{
 		if (buildstate->rowstoskip < 0)
-			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, list_length(buildstate->samples), targsamples);
+			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, samples->length, targsamples);
 
 		if (buildstate->rowstoskip <= 0)
 		{
-			ListCell   *lc;
-
 #if PG_VERSION_NUM >= 150000
 			int			k = (int) (targsamples * sampler_random_fract(&buildstate->rstate.randstate));
 #else
@@ -113,8 +100,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 #endif
 
 			Assert(k >= 0 && k < targsamples);
-			lc = list_nth_cell(buildstate->samples, k);
+			VectorArraySet(samples, k, DatumGetVector(value));
-			lfirst(lc) = DatumGetVector(value);
 		}
 
 		buildstate->rowstoskip -= 1;
@@ -129,13 +115,21 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 			   bool *isnull, bool tupleIsAlive, void *state)
 {
 	IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
+	MemoryContext oldCtx;
 
 	/* Skip nulls */
 	if (isnull[0])
 		return;
 
+	/* Use memory context since detoast can allocate */
+	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
+
 	/* Add sample */
-	AddSample(values, buildstate);
+	AddSample(values, state);
+
+	/* Reset memory context */
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextReset(buildstate->tmpCtx);
 }
 
 /*
@@ -144,7 +138,7 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 static void
 SampleRows(IvfflatBuildState * buildstate)
 {
-	int			targsamples = buildstate->targsamples;
+	int			targsamples = buildstate->samples->maxlen;
 	BlockNumber totalblocks = RelationGetNumberOfBlocks(buildstate->heap);
 
 	buildstate->rowstoskip = -1;
@@ -455,13 +449,12 @@ ComputeCenters(IvfflatBuildState * buildstate)
 
 	/* Sample rows */
 	/* TODO Ensure within maintenance_work_mem */
-	buildstate->samples = NIL;
+	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions);
-	buildstate->targsamples = numSamples;
 	if (buildstate->heap != NULL)
 	{
 		SampleRows(buildstate);
 
-		if (list_length(buildstate->samples) < buildstate->lists)
+		if (buildstate->samples->length < buildstate->lists)
 		{
 			ereport(NOTICE,
 					(errmsg("ivfflat index created with little data"),
@@ -474,7 +467,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers));
 
 	/* Free samples before we allocate more memory */
-	list_free_deep(buildstate->samples);
+	VectorArrayFree(buildstate->samples);
 }
 
 /*
@@ -550,10 +543,10 @@ 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

@@ -76,8 +76,8 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	List	   *qinfos;
 #endif
 
-	/* Never use index without order */
+	/* Never use index without order or limit */
-	if (path->indexorderbys == NULL)
+	if (path->indexorderbys == NULL || root->limit_tuples < 0)
 	{
 		*indexStartupCost = DBL_MAX;
 		*indexTotalCost = DBL_MAX;
@@ -105,6 +105,20 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
 	 */
 	costs.numIndexTuples = path->indexinfo->tuples * ratio;
 
+	/*
+	 * Do not use index if limit + offset > expected tuples unless
+	 * enable_seqscan = off
+	 */
+	if (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

src/ivfflat.h

@@ -80,10 +80,6 @@
 #define RandomInt() random()
 #endif
 
-#if PG_VERSION_NUM < 130000
-#define list_sort(list, cmp) list_qsort(list, cmp)
-#endif
-
 /* Variables */
 extern int	ivfflat_probes;
 
@@ -182,8 +178,7 @@ typedef struct IvfflatBuildState
 	Oid			collation;
 
 	/* Variables */
-	List	   *samples;
+	VectorArray samples;
-	int			targsamples;
 	VectorArray centers;
 	ListInfo   *listInfo;
 	Vector	   *normvec;
@@ -279,7 +274,7 @@ typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 VectorArray VectorArrayInit(int maxlen, int dimensions);
 void		VectorArrayFree(VectorArray arr);
 void		PrintVectorArray(char *msg, VectorArray arr);
-void		IvfflatKmeans(Relation index, List *samples, VectorArray centers);
+void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 int			IvfflatGetLists(Relation index);

src/ivfkmeans.c

@@ -6,26 +6,30 @@
 #include "ivfflat.h"
 #include "miscadmin.h"
 
+#ifdef IVFFLAT_MEMORY
+#include "utils/memutils.h"
+#endif
+
 /*
  * Initialize with kmeans++
  *
  * https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf
  */
 static void
-InitCenters(Relation index, List *samples, VectorArray centers, float *lowerBound)
+InitCenters(Relation index, VectorArray samples, VectorArray centers, float *lowerBound)
 {
 	FmgrInfo   *procinfo;
 	Oid			collation;
 	int64		j;
-	float	   *weight = palloc(list_length(samples) * sizeof(float));
+	float	   *weight = palloc(samples->length * sizeof(float));
 	int			numCenters = centers->maxlen;
-	int			numSamples = list_length(samples);
+	int			numSamples = samples->length;
 
 	procinfo = index_getprocinfo(index, 1, IVFFLAT_KMEANS_DISTANCE_PROC);
 	collation = index->rd_indcollation[0];
 
 	/* Choose an initial center uniformly at random */
-	VectorArraySet(centers, 0, list_nth(samples, RandomInt() % list_length(samples)));
+	VectorArraySet(centers, 0, VectorArrayGet(samples, RandomInt() % samples->length));
 	centers->length++;
 
 	for (j = 0; j < numSamples; j++)
@@ -42,7 +46,7 @@ InitCenters(Relation index, List *samples, VectorArray centers, float *lowerBoun
 
 		for (j = 0; j < numSamples; j++)
 		{
-			Vector	   *vec = list_nth(samples, j);
+			Vector	   *vec = VectorArrayGet(samples, j);
 			double		distance;
 
 			/* Only need to compute distance for new center */
@@ -74,7 +78,7 @@ InitCenters(Relation index, List *samples, VectorArray centers, float *lowerBoun
 				break;
 		}
 
-		VectorArraySet(centers, i + 1, list_nth(samples, j));
+		VectorArraySet(centers, i + 1, VectorArrayGet(samples, j));
 		centers->length++;
 	}
 
@@ -106,41 +110,25 @@ CompareVectors(const void *a, const void *b)
 	return vector_cmp_internal((Vector *) a, (Vector *) b);
 }
 
-/*
- * Compare list vectors
- */
-static int
-#if PG_VERSION_NUM >= 130000
-CompareListVectors(const ListCell *a, const ListCell *b)
-#else
-CompareListVectors(const void *a, const void *b)
-#endif
-{
-	Vector	   *va = lfirst((ListCell *) a);
-	Vector	   *vb = lfirst((ListCell *) b);
-
-	return CompareVectors(va, vb);
-}
-
 /*
  * Quick approach if we have little data
  */
 static void
-QuickCenters(Relation index, List *samples, VectorArray centers)
+QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 {
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 
 	/* Copy existing vectors while avoiding duplicates */
-	if (list_length(samples) > 0)
+	if (samples->length > 0)
 	{
-		list_sort(samples, CompareListVectors);
+		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
-		for (int i = 0; i < list_length(samples); i++)
+		for (int i = 0; i < samples->length; i++)
 		{
-			Vector	   *vec = list_nth(samples, i);
+			Vector	   *vec = VectorArrayGet(samples, i);
 
-			if (i == 0 || CompareVectors(vec, list_nth(samples, i - 1)) != 0)
+			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
 			{
 				VectorArraySet(centers, centers->length, vec);
 				centers->length++;
@@ -167,6 +155,23 @@ QuickCenters(Relation index, List *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.
  *
@@ -176,7 +181,7 @@ QuickCenters(Relation index, List *samples, VectorArray centers)
  * https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf
  */
 static void
-ElkanKmeans(Relation index, List *samples, VectorArray centers)
+ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
@@ -187,7 +192,7 @@ ElkanKmeans(Relation index, List *samples, VectorArray centers)
 	int64		k;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
-	int			numSamples = list_length(samples);
+	int			numSamples = samples->length;
 	VectorArray newCenters;
 	int		   *centerCounts;
 	int		   *closestCenters;
@@ -198,7 +203,7 @@ ElkanKmeans(Relation index, List *samples, VectorArray centers)
 	float	   *newcdist;
 
 	/* Calculate allocation sizes */
-	Size		samplesSize = 0;
+	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
 	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->dim);
 	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, dimensions);
 	Size		centerCountsSize = sizeof(int) * numCenters;
@@ -247,6 +252,10 @@ ElkanKmeans(Relation index, List *samples, VectorArray centers)
 		vec->dim = dimensions;
 	}
 
+#ifdef IVFFLAT_MEMORY
+	ShowMemoryUsage(totalSize);
+#endif
+
 	/* Pick initial centers */
 	InitCenters(index, samples, centers, lowerBound);
 
@@ -342,7 +351,7 @@ ElkanKmeans(Relation index, List *samples, VectorArray centers)
 				if (upperBound[j] <= halfcdist[closestCenters[j] * numCenters + k])
 					continue;
 
-				vec = list_nth(samples, j);
+				vec = VectorArrayGet(samples, j);
 
 				/* Step 3a */
 				if (rj)
@@ -393,7 +402,7 @@ ElkanKmeans(Relation index, List *samples, VectorArray centers)
 		{
 			int			closestCenter;
 
-			vec = list_nth(samples, j);
+			vec = VectorArrayGet(samples, j);
 			closestCenter = closestCenters[j];
 
 			/* Increment sum and count of closest center */
@@ -530,9 +539,9 @@ CheckCenters(Relation index, VectorArray centers)
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, List *samples, VectorArray centers)
+IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers)
 {
-	if (list_length(samples) <= centers->maxlen)
+	if (samples->length <= centers->maxlen)
 		QuickCenters(index, samples, centers);
 	else
 		ElkanKmeans(index, samples, centers);

src/vector.c

@@ -89,7 +89,7 @@ CheckDim(int dim)
 }
 
 /*
- * Ensure finite elements
+ * Ensure finite element
  */
 static inline void
 CheckElement(float value)
@@ -177,14 +177,15 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_in);
 Datum
 vector_in(PG_FUNCTION_ARGS)
 {
-	char	   *str = PG_GETARG_CSTRING(0);
+	char	   *lit = 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	   *lit = pstrdup(str);
+	char	   *litcopy = pstrdup(lit);
+	char	   *str = litcopy;
 
 	while (vector_isspace(*str))
 		str++;
@@ -268,7 +269,7 @@ vector_in(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_DATA_EXCEPTION),
 				 errmsg("vector must have at least 1 dimension")));
 
-	pfree(lit);
+	pfree(litcopy);
 
 	CheckExpectedDim(typmod, dim);
 
@@ -437,17 +438,18 @@ 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	   *arg = PG_GETARG_VECTOR_P(0);
+	Vector	   *vec = PG_GETARG_VECTOR_P(0);
 	int32		typmod = PG_GETARG_INT32(1);
 
-	CheckExpectedDim(typmod, arg->dim);
+	CheckExpectedDim(typmod, vec->dim);
 
-	PG_RETURN_POINTER(arg);
+	PG_RETURN_POINTER(vec);
 }
 
 /*
@@ -464,7 +466,6 @@ array_to_vector(PG_FUNCTION_ARGS)
 	bool		typbyval;
 	char		typalign;
 	Datum	   *elemsp;
-	bool	   *nullsp;
 	int			nelemsp;
 
 	if (ARR_NDIM(array) > 1)
@@ -478,7 +479,7 @@ 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, &nullsp, &nelemsp);
+	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, NULL, &nelemsp);
 
 	CheckDim(nelemsp);
 	CheckExpectedDim(typmod, nelemsp);
@@ -512,6 +513,12 @@ 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]);
@@ -718,7 +725,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 }
 
 /*
- * Get the L1 distance between vectors
+ * Get the L1 distance between two vectors
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(l1_distance);
 Datum
@@ -897,8 +904,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_lt);
 Datum
 vector_lt(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) < 0);
 }
@@ -910,8 +917,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_le);
 Datum
 vector_le(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) <= 0);
 }
@@ -923,8 +930,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_eq);
 Datum
 vector_eq(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) == 0);
 }
@@ -936,8 +943,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ne);
 Datum
 vector_ne(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) != 0);
 }
@@ -949,8 +956,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ge);
 Datum
 vector_ge(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) >= 0);
 }
@@ -962,8 +969,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_gt);
 Datum
 vector_gt(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_BOOL(vector_cmp_internal(a, b) > 0);
 }
@@ -975,8 +982,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_cmp);
 Datum
 vector_cmp(PG_FUNCTION_ARGS)
 {
-	Vector	   *a = (Vector *) PG_GETARG_VECTOR_P(0);
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
-	Vector	   *b = (Vector *) PG_GETARG_VECTOR_P(1);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
 
 	PG_RETURN_INT32(vector_cmp_internal(a, b));
 }

test/expected/hnsw_cosine.out

@@ -3,7 +3,7 @@ CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t USING hnsw (val vector_cosine_ops);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+SELECT * FROM t ORDER BY val <=> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,1,1]
@@ -11,13 +11,13 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,2,4]
 (3 rows)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]' LIMIT 5) t2;
  count 
 -------
      3
 (1 row)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector) LIMIT 5) t2;
  count 
 -------
      3

test/expected/hnsw_ip.out

@@ -3,7 +3,7 @@ CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t USING hnsw (val vector_ip_ops);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <#> '[3,3,3]';
+SELECT * FROM t ORDER BY val <#> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,4]
@@ -12,7 +12,7 @@ SELECT * FROM t ORDER BY val <#> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector) LIMIT 5) t2;
  count 
 -------
      4

test/expected/hnsw_l2.out

@@ -3,7 +3,7 @@ CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,3]
@@ -12,7 +12,7 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector) LIMIT 5;
    val   
 ---------
  [0,0,0]
@@ -28,7 +28,7 @@ SELECT COUNT(*) FROM t;
 (1 row)
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
  val 
 -----
 (0 rows)

test/expected/hnsw_unlogged.out

@@ -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 hnsw (val vector_l2_ops);
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,3]

test/expected/ivfflat_cosine.out

@@ -3,7 +3,7 @@ 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_cosine_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+SELECT * FROM t ORDER BY val <=> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,1,1]
@@ -11,13 +11,13 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,2,4]
 (3 rows)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]' LIMIT 5) t2;
  count 
 -------
      3
 (1 row)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector) LIMIT 5) t2;
  count 
 -------
      3

test/expected/ivfflat_ip.out

@@ -3,7 +3,7 @@ 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_ip_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <#> '[3,3,3]';
+SELECT * FROM t ORDER BY val <#> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,4]
@@ -12,7 +12,7 @@ SELECT * FROM t ORDER BY val <#> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector) LIMIT 5) t2;
  count 
 -------
      4

test/expected/ivfflat_l2.out

@@ -3,7 +3,7 @@ CREATE TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,3]
@@ -12,7 +12,7 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector) LIMIT 5;
    val   
 ---------
  [0,0,0]
@@ -31,7 +31,7 @@ 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]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
  val 
 -----
 (0 rows)

test/expected/ivfflat_unlogged.out

@@ -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);
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
    val   
 ---------
  [1,2,3]

test/sql/hnsw_cosine.sql

@@ -6,8 +6,8 @@ CREATE INDEX ON t USING hnsw (val vector_cosine_ops);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+SELECT * FROM t ORDER BY val <=> '[3,3,3]' LIMIT 5;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]' LIMIT 5) t2;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector) LIMIT 5) t2;
 
 DROP TABLE t;

test/sql/hnsw_ip.sql

@@ -6,7 +6,7 @@ CREATE INDEX ON t USING hnsw (val vector_ip_ops);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
-SELECT * FROM t ORDER BY val <#> '[3,3,3]';
+SELECT * FROM t ORDER BY val <#> '[3,3,3]' LIMIT 5;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector) LIMIT 5) t2;
 
 DROP TABLE t;

test/sql/hnsw_l2.sql

@@ -6,11 +6,11 @@ CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector) LIMIT 5;
 SELECT COUNT(*) FROM t;
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
 
 DROP TABLE t;

test/sql/hnsw_unlogged.sql

@@ -4,6 +4,6 @@ CREATE UNLOGGED TABLE t (val vector(3));
 INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
 
 DROP TABLE t;

test/sql/ivfflat_cosine.sql

@@ -6,8 +6,8 @@ CREATE INDEX ON t USING ivfflat (val vector_cosine_ops) WITH (lists = 1);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
+SELECT * FROM t ORDER BY val <=> '[3,3,3]' LIMIT 5;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]' LIMIT 5) t2;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector) LIMIT 5) t2;
 
 DROP TABLE t;

test/sql/ivfflat_ip.sql

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

test/sql/ivfflat_l2.sql

@@ -6,11 +6,11 @@ CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
 
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
-SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector) LIMIT 5;
 SELECT COUNT(*) FROM t;
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
 
 DROP TABLE t;

test/sql/ivfflat_unlogged.sql

@@ -4,6 +4,6 @@ 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);
 
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
+SELECT * FROM t ORDER BY val <-> '[3,3,3]' LIMIT 5;
 
 DROP TABLE t;

test/t/007_ivfflat_inserts.pl

@@ -49,7 +49,7 @@ is(idx_scan(), 0);
 $count = $node->safe_psql("postgres", qq(
 	SET enable_seqscan = off;
 	SET ivfflat.probes = 100;
-	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
+	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 20000) t;
 ));
 is($count, $expected);
 is(idx_scan(), 1);

test/t/012_hnsw_build_recall.pl

@@ -83,11 +83,32 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Add index
+	# Build index serially
-	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
+	$node->safe_psql("postgres", qq(
+		SET max_parallel_maintenance_workers = 0;
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
 
+	# Test approximate results
 	my $min = $operator eq "<#>" ? 0.80 : 0.99;
 	test_recall($min, $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/014_hnsw_inserts.pl

@@ -42,7 +42,7 @@ for my $i (1 .. 20)
 
 	my $count = $node->safe_psql("postgres", qq(
 		SET enable_seqscan = off;
-		SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
+		SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 20) t;
 	));
 	is($count, 10);
 
@@ -63,7 +63,7 @@ $node->pgbench(
 my $count = $node->safe_psql("postgres", qq(
 	SET enable_seqscan = off;
 	SET hnsw.ef_search = 1000;
-	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1)) t;
+	SELECT COUNT(*) FROM (SELECT v FROM tst ORDER BY v <-> (SELECT v FROM tst LIMIT 1) LIMIT 1000) t;
 ));
 # Elements may lose all incoming connections with the HNSW algorithm
 # Vacuuming can fix this if one of the elements neighbors is deleted

test/t/015_hnsw_duplicates.pl

@@ -26,7 +26,7 @@ sub test_duplicates
 	my $res = $node->safe_psql("postgres", qq(
 		SET enable_seqscan = off;
 		SET hnsw.ef_search = 1;
-		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]') t;
+		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]' LIMIT 20) t;
 	));
 	is($res, 10);
 }

test/t/019_ivfflat_limit.pl

@@ -0,0 +1,64 @@
+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

@@ -0,0 +1,62 @@
+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();