Use List for samples

.github/workflows/build.yml

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

CHANGELOG.md

@@ -1,7 +1,3 @@
-## 0.5.2 (unreleased)
-
-- Added support for on-disk parallel index builds for HNSW
-
 ## 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)"
-	for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+	copy $(HEADERS) "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 installcheck:
 	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)

README.md

@@ -26,7 +26,7 @@ make install # may need sudo
 
 See the [installation notes](#installation-notes) if you run into issues
 
-You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
+You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres)
 
 ## Getting Started
 
@@ -215,23 +215,6 @@ 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, round(100.0 * tuples_done / nullif(tuples_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
-```
-
-The phases for IVFFlat are:
-
-1. `initializing`
-2. `performing k-means`
-3. `assigning tuples`
-4. `loading tuples`
-
-Note: `%` is only populated during the `loading tuples` phase
-
 ## HNSW
 
 An HNSW index creates a multilayer graph. It has slower build times and uses more memory than IVFFlat, but has better query performance (in terms of speed-recall tradeoff). There’s no training step like IVFFlat, so the index can be created without any data in the table.
@@ -288,18 +271,22 @@ SELECT ...
 COMMIT;
 ```
 
-### Indexing Progress
+## 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;
+SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
 ```
 
-The phases for HNSW are:
+The phases are:
 
 1. `initializing`
-2. `loading tuples`
+2. `performing k-means` - IVFFlat only
+3. `assigning tuples` - IVFFlat only
+4. `loading tuples`
+
+Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
 
 ## Filtering
 
@@ -330,15 +317,13 @@ CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(cate
 
 ## Hybrid Search
 
-Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search.
+Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search ([Python example](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py)).
 
 ```sql
 SELECT id, content FROM items, plainto_tsquery('hello search') query
     WHERE textsearch @@ query ORDER BY ts_rank_cd(textsearch, query) DESC LIMIT 5;
 ```
 
-You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
-
 ## Performance
 
 Use `EXPLAIN ANALYZE` to debug performance.
@@ -375,7 +360,6 @@ 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)
 Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
@@ -383,11 +367,10 @@ Dart | [pgvector-dart](https://github.com/pgvector/pgvector-dart)
 Elixir | [pgvector-elixir](https://github.com/pgvector/pgvector-elixir)
 Go | [pgvector-go](https://github.com/pgvector/pgvector-go)
 Haskell | [pgvector-haskell](https://github.com/pgvector/pgvector-haskell)
-Java, Kotlin, Groovy, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
-JavaScript, TypeScript | [pgvector-node](https://github.com/pgvector/pgvector-node)
+Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
 Julia | [pgvector-julia](https://github.com/pgvector/pgvector-julia)
 Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
-Nim | [pgvector-nim](https://github.com/pgvector/pgvector-nim)
+Node.js | [pgvector-node](https://github.com/pgvector/pgvector-node)
 Perl | [pgvector-perl](https://github.com/pgvector/pgvector-perl)
 PHP | [pgvector-php](https://github.com/pgvector/pgvector-php)
 Python | [pgvector-python](https://github.com/pgvector/pgvector-python)
@@ -395,7 +378,6 @@ R | [pgvector-r](https://github.com/pgvector/pgvector-r)
 Ruby | [pgvector-ruby](https://github.com/pgvector/pgvector-ruby), [Neighbor](https://github.com/ankane/neighbor)
 Rust | [pgvector-rust](https://github.com/pgvector/pgvector-rust)
 Swift | [pgvector-swift](https://github.com/pgvector/pgvector-swift)
-Zig | [pgvector-zig](https://github.com/pgvector/pgvector-zig)
 
 ## Frequently Asked Questions
 
@@ -411,55 +393,6 @@ Yes, pgvector uses the write-ahead log (WAL), which allows for replication and p
 
 You’ll need to use [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction) at the moment.
 
-#### Can I store vectors with different dimensions in the same column?
-
-You can use `vector` as the type (instead of `vector(3)`).
-
-```sql
-CREATE TABLE embeddings (model_id bigint, item_id bigint, embedding vector, PRIMARY KEY (model_id, item_id));
-```
-
-However, you can only create indexes on rows with the same number of dimensions (using [expression](https://www.postgresql.org/docs/current/indexes-expressional.html) and [partial](https://www.postgresql.org/docs/current/indexes-partial.html) indexing):
-
-```sql
-CREATE INDEX ON embeddings USING hnsw ((embedding::vector(3)) vector_l2_ops) WHERE (model_id = 123);
-```
-
-and query with:
-
-```sql
-SELECT * FROM embeddings WHERE model_id = 123 ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
-```
-
-#### Can I store vectors with more precision?
-
-You can use the `double precision[]` or `numeric[]` type to store vectors with more precision.
-
-```sql
-CREATE TABLE items (id bigserial PRIMARY KEY, embedding double precision[]);
-
--- use {} instead of [] for Postgres arrays
-INSERT INTO items (embedding) VALUES ('{1,2,3}'), ('{4,5,6}');
-```
-
-Optionally, add a [check constraint](https://www.postgresql.org/docs/current/ddl-constraints.html) to ensure data can be converted to the `vector` type and has the expected dimensions.
-
-```sql
-ALTER TABLE items ADD CHECK (vector_dims(embedding::vector) = 3);
-```
-
-Use [expression indexing](https://www.postgresql.org/docs/current/indexes-expressional.html) to index (at a lower precision):
-
-```sql
-CREATE INDEX ON items USING hnsw ((embedding::vector(3)) vector_l2_ops);
-```
-
-and query with:
-
-```sql
-SELECT * FROM items ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
-```
-
 ## Troubleshooting
 
 #### Why isn’t a query using an index?
@@ -473,8 +406,6 @@ SELECT ...
 COMMIT;
 ```
 
-Also, if the table is small, a table scan may be faster.
-
 #### Why isn’t a query using a parallel table scan?
 
 The planner doesn’t consider [out-of-line storage](https://www.postgresql.org/docs/current/storage-toast.html) in cost estimates, which can make a serial scan look cheaper. You can reduce the cost of a parallel scan for a query with:
@@ -664,7 +595,7 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
 
 ## Upgrading
 
-[Install](#installation) the latest version (use the same method as the original installation). Then in each database you want to upgrade, run:
+Install the latest version. Then in each database you want to upgrade, run:
 
 ```sql
 ALTER EXTENSION vector UPDATE;
@@ -678,16 +609,6 @@ SELECT extversion FROM pg_extension WHERE extname = 'vector';
 
 ## Upgrade Notes
 
-### 0.6.0
-
-If upgrading with Postgres < 13, remove this line from `sql/vector--0.5.1--0.6.0.sql`:
-
-```sql
-ALTER TYPE vector SET (STORAGE = external);
-```
-
-Then run `make install` and `ALTER EXTENSION vector UPDATE;`.
-
 ### 0.4.0
 
 If upgrading with Postgres < 13, remove this line from `sql/vector--0.3.2--0.4.0.sql`:

sql/vector--0.5.1--0.6.0.sql

@@ -1,5 +0,0 @@
--- complain if script is sourced in psql, rather than via CREATE EXTENSION
-\echo Use "ALTER EXTENSION vector UPDATE TO '0.6.0'" to load this file. \quit
-
--- remove this single line for Postgres < 13
-ALTER TYPE vector SET (STORAGE = external);

sql/vector.sql

@@ -26,7 +26,7 @@ CREATE TYPE vector (
 	TYPMOD_IN = vector_typmod_in,
 	RECEIVE   = vector_recv,
 	SEND      = vector_send,
-	STORAGE   = external
+	STORAGE   = extended
 );
 
 -- functions

src/hnsw.c

@@ -14,7 +14,6 @@
 #endif
 
 int			hnsw_ef_search;
-bool		hnsw_enable_parallel_build;
 static relopt_kind hnsw_relopt_kind;
 
 /*
@@ -40,11 +39,6 @@ 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);
 }
 
 /*

src/hnsw.h

@@ -4,7 +4,6 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
-#include "access/parallel.h"
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
@@ -15,10 +14,6 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
-#if PG_VERSION_NUM < 120000
-#include "access/relscan.h"
-#endif
-
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
@@ -64,7 +59,7 @@
 
 #define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
 
-#define HNSW_ELEMENT_TUPLE_SIZE(size)	MAXALIGN(offsetof(HnswElementTupleData, data) + (size))
+#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -95,7 +90,6 @@
 
 /* Variables */
 extern int	hnsw_ef_search;
-extern bool hnsw_enable_parallel_build;
 
 typedef struct HnswNeighborArray HnswNeighborArray;
 
@@ -109,7 +103,7 @@ typedef struct HnswElementData
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Datum		value;
+	Vector	   *vec;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -142,49 +136,6 @@ typedef struct HnswOptions
 	int			efConstruction; /* size of dynamic candidate list */
 }			HnswOptions;
 
-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;
-	double		indtuples;
-
-#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 */
@@ -212,16 +163,12 @@ typedef struct HnswBuildState
 	HnswElement entryPoint;
 	double		ml;
 	int			maxLevel;
-	long		memoryLeft;
+	double		maxInMemoryElements;
 	bool		flushed;
 	Vector	   *normvec;
 
 	/* Memory */
 	MemoryContext tmpCtx;
-
-	/* Parallel builds */
-	HnswLeader *hnswleader;
-	HnswShared *hnswshared;
 }			HnswBuildState;
 
 typedef struct HnswMetaPageData
@@ -257,7 +204,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		data;
+	Vector		vec;
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -342,7 +289,6 @@ void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation i
 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);

src/hnswbuild.c

@@ -2,16 +2,12 @@
 
 #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
@@ -39,23 +35,6 @@
 #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)
-
 /*
  * Create the metapage
  */
@@ -126,7 +105,8 @@ CreateElementPages(HnswBuildState * buildstate)
 {
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
-	Size		etupAllocSize;
+	int			dimensions = buildstate->dimensions;
+	Size		etupSize;
 	Size		maxSize;
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
@@ -137,11 +117,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	ListCell   *lc;
 
 	/* Calculate sizes */
-	etupAllocSize = BLCKSZ;
 	maxSize = HNSW_MAX_SIZE;
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
-	etup = palloc0(etupAllocSize);
+	etup = palloc0(etupSize);
 	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
@@ -153,24 +133,15 @@ CreateElementPages(HnswBuildState * buildstate)
 	foreach(lc, buildstate->elements)
 	{
 		HnswElement element = lfirst(lc);
-		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
 
-		/* Zero memory for each element */
-		MemSet(etup, 0, etupAllocSize);
+		HnswSetElementTuple(etup, element);
 
 		/* 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);
@@ -293,14 +264,13 @@ FlushPages(HnswBuildState * buildstate)
  * Insert tuple
  */
 static bool
-InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup, MemoryContext outerCtx)
+InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup)
 {
 	FmgrInfo   *procinfo = buildstate->procinfo;
 	Oid			collation = buildstate->collation;
 	HnswElement entryPoint = buildstate->entryPoint;
 	int			efConstruction = buildstate->efConstruction;
 	int			m = buildstate->m;
-	MemoryContext oldCtx;
 
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -313,9 +283,7 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	}
 
 	/* Copy value to element so accessible outside of memory context */
-	oldCtx = MemoryContextSwitchTo(outerCtx);
-	element->value = datumCopy(value, false, -1);
-	MemoryContextSwitchTo(oldCtx);
+	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
 
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
@@ -345,21 +313,6 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	return *dup == NULL;
 }
 
-/*
- * 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 += sizeof(ItemPointerData);
-	elementSize += VARSIZE_ANY(DatumGetPointer(e->value));
-	return elementSize;
-}
-
 /*
  * Callback for table_index_build_scan
  */
@@ -381,7 +334,7 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (isnull[0])
 		return;
 
-	if (buildstate->memoryLeft <= 0)
+	if (buildstate->indtuples >= buildstate->maxInMemoryElements)
 	{
 		if (!buildstate->flushed)
 		{
@@ -396,18 +349,7 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
-		{
-			if (buildstate->hnswshared)
-			{
-				HnswShared *hnswshared = buildstate->hnswshared;
-
-				SpinLockAcquire(&hnswshared->mutex);
-				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++hnswshared->indtuples);
-				SpinLockRelease(&hnswshared->mutex);
-			}
-			else
-				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
-		}
+			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
 
 		/* Reset memory context */
 		MemoryContextSwitchTo(oldCtx);
@@ -418,12 +360,13 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* 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, oldCtx);
+	inserted = InsertTuple(index, values, element, buildstate, &dup);
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -431,21 +374,31 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Add outside memory context */
 	if (dup != NULL)
-	{
 		HnswAddHeapTid(dup, tid);
-		buildstate->memoryLeft -= sizeof(ItemPointerData);
-	}
 
 	/* Add to buildstate or free */
 	if (inserted)
-	{
 		buildstate->elements = lappend(buildstate->elements, element);
-		buildstate->memoryLeft -= HnswElementMemory(element, buildstate->m);
-	}
 	else
 		HnswFreeElement(element);
 }
 
+/*
+ * Get the max number of elements that fit into maintenance_work_mem
+ */
+static double
+HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
+{
+	Size		elementSize = sizeof(HnswElementData);
+	double		avgLevel = -log(0.5) * ml;
+
+	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
+	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
+	elementSize += sizeof(ItemPointerData);
+	elementSize += VECTOR_SIZE(dimensions);
+	return (maintenance_work_mem * 1024L) / elementSize;
+}
+
 /*
  * Initialize the build state
  */
@@ -483,7 +436,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->entryPoint = NULL;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->memoryLeft = maintenance_work_mem * 1024L;
+	buildstate->maxInMemoryElements = HnswGetMaxInMemoryElements(buildstate->m, buildstate->ml, buildstate->dimensions);
 	buildstate->flushed = false;
 
 	/* Reuse for each tuple */
@@ -492,9 +445,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
-
-	buildstate->hnswleader = NULL;
-	buildstate->hnswshared = NULL;
 }
 
 /*
@@ -507,374 +457,21 @@ FreeBuildState(HnswBuildState * buildstate)
 	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->indtuples = hnswshared->indtuples;
-			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);
-	/* TODO Support in-memory builds */
-	buildstate.memoryLeft = 0;
-	buildstate.flushed = true;
-	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;
-	hnswshared->indtuples = 0;
-#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);
-}
-
 /*
  * 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 (hnsw_enable_parallel_build)
-		parallel_workers = plan_create_index_workers(RelationGetRelid(buildstate->heap), RelationGetRelid(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 sort */
-	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,
-													   true, true, BuildCallback, (void *) buildstate, NULL);
+	buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+												   true, true, BuildCallback, (void *) buildstate, NULL);
 #else
-		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, BuildCallback, (void *) buildstate, NULL);
+	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+											   true, BuildCallback, (void *) buildstate, NULL);
 #endif
-	}
-
-	/* End parallel build */
-	if (buildstate->hnswleader)
-		HnswEndParallel(buildstate->hnswleader);
 }
 
 /*

src/hnswinsert.c

@@ -123,6 +123,7 @@ 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;
@@ -131,7 +132,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -404,9 +405,8 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	ItemId		itemid;
+	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
 	HnswElementTuple etup;
-	Size		etupSize;
 	int			i;
 
 	/* Read page */
@@ -416,9 +416,7 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	page = GenericXLogRegisterBuffer(state, buf, 0);
 
 	/* Find space */
-	itemid = PageGetItemId(page, dup->offno);
-	etup = (HnswElementTuple) PageGetItem(page, itemid);
-	etupSize = ItemIdGetLength(itemid);
+	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
 	for (i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (!ItemPointerIsValid(&etup->heaptids[i]))
@@ -517,7 +515,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 
 	/* Create an element */
 	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->value = value;
+	element->vec = DatumGetVector(value);
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)

src/hnswutils.c

@@ -4,7 +4,6 @@
 
 #include "hnsw.h"
 #include "storage/bufmgr.h"
-#include "utils/datum.h"
 #include "vector.h"
 
 /*
@@ -188,8 +187,7 @@ HnswFreeElement(HnswElement element)
 {
 	HnswFreeNeighbors(element);
 	list_free_deep(element->heaptids);
-	if (DatumGetPointer(element->value))
-		pfree(DatumGetPointer(element->value));
+	pfree(element->vec);
 	pfree(element);
 }
 
@@ -216,7 +214,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->value = PointerGetDatum(NULL);
+	element->vec = NULL;
 	return element;
 }
 
@@ -326,7 +324,7 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->data, DatumGetPointer(element->value), VARSIZE_ANY(DatumGetPointer(element->value)));
+	memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
 }
 
 /*
@@ -448,7 +446,10 @@ 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));
+	}
 }
 
 /*
@@ -475,7 +476,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
+		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
 
 	UnlockReleaseBuffer(buf);
 }
@@ -486,7 +487,7 @@ 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, hc->element->value));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
 }
 
 /*
@@ -749,7 +750,7 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
 		}
 	}
 
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
 }
 
 /*
@@ -876,7 +877,7 @@ HnswFindDuplicate(HnswElement e)
 		HnswCandidate *neighbor = &neighbors->items[i];
 
 		/* Exit early since ordered by distance */
-		if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
+		if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
 			break;
 
 		/* Check for space */
@@ -929,13 +930,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = hc->element->value;
+			Datum		q = PointerGetDatum(hc->element->vec);
 
 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (DatumGetPointer(hc3->element->value) == NULL)
+				if (hc3->element->vec == NULL)
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -1016,7 +1017,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = element->value;
+	Datum		q = PointerGetDatum(element->vec);
 	HnswElement skipElement = existing ? element : NULL;
 
 	/* No neighbors if no entry point */

src/hnswvacuum.c

@@ -62,8 +62,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		/* Iterate over nodes */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			ItemId		itemid = PageGetItemId(page, offno);
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
 			int			idx = 0;
 			bool		itemUpdated = false;
 
@@ -94,7 +93,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 				if (itemUpdated)
 				{
-					Size		etupSize = ItemIdGetLength(itemid);
+					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -478,8 +477,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		/* Update element and neighbors together */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			ItemId		itemid = PageGetItemId(page, offno);
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
@@ -507,7 +505,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				continue;
 
 			/* Calculate sizes */
-			etupSize = ItemIdGetLength(itemid);
+			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -530,7 +528,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
+			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)

src/ivfbuild.c

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

src/ivfflat.h

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

@@ -12,20 +12,20 @@
  * https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf
  */
 static void
-InitCenters(Relation index, VectorArray samples, VectorArray centers, float *lowerBound)
+InitCenters(Relation index, List *samples, VectorArray centers, float *lowerBound)
 {
 	FmgrInfo   *procinfo;
 	Oid			collation;
 	int64		j;
-	float	   *weight = palloc(samples->length * sizeof(float));
+	float	   *weight = palloc(list_length(samples) * sizeof(float));
 	int			numCenters = centers->maxlen;
-	int			numSamples = samples->length;
+	int			numSamples = list_length(samples);
 
 	procinfo = index_getprocinfo(index, 1, IVFFLAT_KMEANS_DISTANCE_PROC);
 	collation = index->rd_indcollation[0];
 
 	/* Choose an initial center uniformly at random */
-	VectorArraySet(centers, 0, VectorArrayGet(samples, RandomInt() % samples->length));
+	VectorArraySet(centers, 0, list_nth(samples, RandomInt() % list_length(samples)));
 	centers->length++;
 
 	for (j = 0; j < numSamples; j++)
@@ -42,7 +42,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 
 		for (j = 0; j < numSamples; j++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, j);
+			Vector	   *vec = list_nth(samples, j);
 			double		distance;
 
 			/* Only need to compute distance for new center */
@@ -74,7 +74,7 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 				break;
 		}
 
-		VectorArraySet(centers, i + 1, VectorArrayGet(samples, j));
+		VectorArraySet(centers, i + 1, list_nth(samples, j));
 		centers->length++;
 	}
 
@@ -106,25 +106,41 @@ 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, VectorArray samples, VectorArray centers)
+QuickCenters(Relation index, List *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 (samples->length > 0)
+	if (list_length(samples) > 0)
 	{
-		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
-		for (int i = 0; i < samples->length; i++)
+		list_sort(samples, CompareListVectors);
+		for (int i = 0; i < list_length(samples); i++)
 		{
-			Vector	   *vec = VectorArrayGet(samples, i);
+			Vector	   *vec = list_nth(samples, i);
 
-			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
+			if (i == 0 || CompareVectors(vec, list_nth(samples, i - 1)) != 0)
 			{
 				VectorArraySet(centers, centers->length, vec);
 				centers->length++;
@@ -160,7 +176,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
  * https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf
  */
 static void
-ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
+ElkanKmeans(Relation index, List *samples, VectorArray centers)
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
@@ -171,7 +187,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	int64		k;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
-	int			numSamples = samples->length;
+	int			numSamples = list_length(samples);
 	VectorArray newCenters;
 	int		   *centerCounts;
 	int		   *closestCenters;
@@ -182,7 +198,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 	float	   *newcdist;
 
 	/* Calculate allocation sizes */
-	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
+	Size		samplesSize = 0;
 	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->dim);
 	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, dimensions);
 	Size		centerCountsSize = sizeof(int) * numCenters;
@@ -326,7 +342,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 				if (upperBound[j] <= halfcdist[closestCenters[j] * numCenters + k])
 					continue;
 
-				vec = VectorArrayGet(samples, j);
+				vec = list_nth(samples, j);
 
 				/* Step 3a */
 				if (rj)
@@ -377,7 +393,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		{
 			int			closestCenter;
 
-			vec = VectorArrayGet(samples, j);
+			vec = list_nth(samples, j);
 			closestCenter = closestCenters[j];
 
 			/* Increment sum and count of closest center */
@@ -514,9 +530,9 @@ CheckCenters(Relation index, VectorArray centers)
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers)
+IvfflatKmeans(Relation index, List *samples, VectorArray centers)
 {
-	if (samples->length <= centers->maxlen)
+	if (list_length(samples) <= centers->maxlen)
 		QuickCenters(index, samples, centers);
 	else
 		ElkanKmeans(index, samples, centers);

src/vector.c

@@ -89,7 +89,7 @@ CheckDim(int dim)
 }
 
 /*
- * Ensure finite element
+ * Ensure finite elements
  */
 static inline void
 CheckElement(float value)
@@ -437,18 +437,17 @@ vector_send(PG_FUNCTION_ARGS)
 
 /*
  * Convert vector to vector
- * This is needed to check the type modifier
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(vector);
 Datum
 vector(PG_FUNCTION_ARGS)
 {
-	Vector	   *vec = PG_GETARG_VECTOR_P(0);
+	Vector	   *arg = PG_GETARG_VECTOR_P(0);
 	int32		typmod = PG_GETARG_INT32(1);
 
-	CheckExpectedDim(typmod, vec->dim);
+	CheckExpectedDim(typmod, arg->dim);
 
-	PG_RETURN_POINTER(vec);
+	PG_RETURN_POINTER(arg);
 }
 
 /*
@@ -465,6 +464,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 	bool		typbyval;
 	char		typalign;
 	Datum	   *elemsp;
+	bool	   *nullsp;
 	int			nelemsp;
 
 	if (ARR_NDIM(array) > 1)
@@ -478,7 +478,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("array must not contain nulls")));
 
 	get_typlenbyvalalign(ARR_ELEMTYPE(array), &typlen, &typbyval, &typalign);
-	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, NULL, &nelemsp);
+	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, &nullsp, &nelemsp);
 
 	CheckDim(nelemsp);
 	CheckExpectedDim(typmod, nelemsp);
@@ -512,12 +512,6 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("unsupported array type")));
 	}
 
-	/*
-	 * Free allocation from deconstruct_array. Do not free individual elements
-	 * when pass-by-reference since they point to original array.
-	 */
-	pfree(elemsp);
-
 	/* Check elements */
 	for (int i = 0; i < result->dim; i++)
 		CheckElement(result->x[i]);

test/t/012_hnsw_build_recall.pl

@@ -83,32 +83,11 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Build index serially
-	$node->safe_psql("postgres", qq(
-		SET max_parallel_maintenance_workers = 0;
-		CREATE INDEX idx ON tst USING hnsw (v $opclass);
-	));
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX 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();