Improved naming [skip ci]

Co-authored-by: Itai Spiegel <itai@mave.com>

.github/workflows/build.yml

@@ -8,10 +8,12 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - postgres: 20
+            os: ubuntu-26.04
           - postgres: 19
-            os: ubuntu-24.04
+            os: ubuntu-26.04
           - postgres: 18
-            os: ubuntu-24.04
+            os: ubuntu-26.04-arm
           - postgres: 17
             os: ubuntu-24.04
           - postgres: 16
@@ -23,7 +25,7 @@ jobs:
           - postgres: 13
             os: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -53,7 +55,7 @@ jobs:
           - postgres: 14
             os: macos-15-intel
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -92,7 +94,7 @@ jobs:
           - postgres: 14
             os: windows-2022
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -133,7 +135,7 @@ jobs:
     if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres-valgrind@v1
         with:
           postgres-version: 18

CHANGELOG.md

@@ -1,3 +1,13 @@
+## 0.8.5 (2026-07-08)
+
+- Reduced memory usage for small tables for IVFFlat index builds
+
+## 0.8.4 (2026-06-30)
+
+- Fixed `hnsw graph not repaired` error with HNSW vacuuming
+- Fixed possible error with inserts during HNSW vacuuming
+- Fixed memory exceeding `maintenance_work_mem` with IVFFlat index builds
+
 ## 0.8.3 (2026-06-17)
 
 - Fixed possible index corruption with HNSW vacuuming

Dockerfile

@@ -5,7 +5,7 @@ ARG DEBIAN_CODENAME=bookworm
 FROM postgres:$PG_MAJOR-$DEBIAN_CODENAME
 ARG PG_MAJOR
 
-ADD https://github.com/pgvector/pgvector.git#v0.8.3 /tmp/pgvector
+ADD https://github.com/pgvector/pgvector.git#v0.8.5 /tmp/pgvector
 
 RUN apt-get update && \
 		apt-mark hold locales && \

META.json

@@ -2,12 +2,12 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.8.3",
+	"version": "0.8.5",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
 	"license": {
-		"PostgreSQL": "http://www.postgresql.org/about/licence"
+		"PostgreSQL": "https://www.postgresql.org/about/licence"
 	},
 	"prereqs": {
 		"runtime": {
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.8.3",
+			"version": "0.8.5",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},
@@ -38,7 +38,7 @@
 	"generated_by": "Andrew Kane",
 	"meta-spec": {
 		"version": "1.0.0",
-		"url": "http://pgxn.org/meta/spec.txt"
+		"url": "https://pgxn.org/meta/spec.txt"
 	},
 	"tags": [
 		"vectors",

Makefile

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.3
+EXTVERSION = 0.8.5
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.3
+EXTVERSION = 0.8.5
 
 DATA_built = sql\$(EXTENSION)--$(EXTVERSION).sql
 OBJS = src\bitutils.obj src\bitvec.obj src\halfutils.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj

README.md

@@ -23,7 +23,7 @@ Compile and install the extension (supports Postgres 13+)
 
 ```sh
 cd /tmp
-git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.5 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -40,7 +40,7 @@ Ensure [C++ support in Visual Studio](https://learn.microsoft.com/en-us/cpp/buil
 ```cmd
 set "PGROOT=C:\Program Files\PostgreSQL\18"
 cd %TEMP%
-git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.5 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -465,6 +465,16 @@ If filtering by many different values, consider [partitioning](https://www.postg
 CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
 ```
 
+## Multitenancy
+
+For applications with multiple tenants, sharing an approximate index between tenants means vectors from one tenant can affect recall (and speed) for other tenants.
+
+For tenant isolation, use [list partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html) or separate tables.
+
+```sql
+CREATE TABLE items (customer_id int, embedding vector(3)) PARTITION BY LIST(customer_id);
+```
+
 ## Iterative Index Scans
 
 With approximate indexes, queries with filtering can return less results since filtering is applied *after* the index is scanned. Starting with 0.8.0, you can enable iterative index scans, which will automatically scan more of the index until enough results are found (or it reaches `hnsw.max_scan_tuples` or `ivfflat.max_probes`).
@@ -1151,23 +1161,23 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 
 Supported tags are:
 
-- `pg18-trixie`, `0.8.3-pg18-trixie`
-- `pg18-bookworm`, `0.8.3-pg18-bookworm`, `pg18`, `0.8.3-pg18`
-- `pg17-trixie`, `0.8.3-pg17-trixie`
-- `pg17-bookworm`, `0.8.3-pg17-bookworm`, `pg17`, `0.8.3-pg17`
-- `pg16-trixie`, `0.8.3-pg16-trixie`
-- `pg16-bookworm`, `0.8.3-pg16-bookworm`, `pg16`, `0.8.3-pg16`
-- `pg15-trixie`, `0.8.3-pg15-trixie`
-- `pg15-bookworm`, `0.8.3-pg15-bookworm`, `pg15`, `0.8.3-pg15`
-- `pg14-trixie`, `0.8.3-pg14-trixie`
-- `pg14-bookworm`, `0.8.3-pg14-bookworm`, `pg14`, `0.8.3-pg14`
-- `pg13-trixie`, `0.8.3-pg13-trixie`
-- `pg13-bookworm`, `0.8.3-pg13-bookworm`, `pg13`, `0.8.3-pg13`
+- `pg18-trixie`, `0.8.5-pg18-trixie`
+- `pg18-bookworm`, `0.8.5-pg18-bookworm`, `pg18`, `0.8.5-pg18`
+- `pg17-trixie`, `0.8.5-pg17-trixie`
+- `pg17-bookworm`, `0.8.5-pg17-bookworm`, `pg17`, `0.8.5-pg17`
+- `pg16-trixie`, `0.8.5-pg16-trixie`
+- `pg16-bookworm`, `0.8.5-pg16-bookworm`, `pg16`, `0.8.5-pg16`
+- `pg15-trixie`, `0.8.5-pg15-trixie`
+- `pg15-bookworm`, `0.8.5-pg15-bookworm`, `pg15`, `0.8.5-pg15`
+- `pg14-trixie`, `0.8.5-pg14-trixie`
+- `pg14-bookworm`, `0.8.5-pg14-bookworm`, `pg14`, `0.8.5-pg14`
+- `pg13-trixie`, `0.8.5-pg13-trixie`
+- `pg13-bookworm`, `0.8.5-pg13-bookworm`, `pg13`, `0.8.5-pg13`
 
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.5 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --pull --build-arg PG_MAJOR=18 -t myuser/pgvector .
 ```

sql/vector--0.8.3--0.8.4.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.8.4'" to load this file. \quit

sql/vector--0.8.4--0.8.5.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.8.5'" to load this file. \quit

src/hnsw.h

@@ -427,7 +427,7 @@ typedef struct HnswVacuumState
 	HnswSupport support;
 
 	/* Variables */
-	struct tidhash_hash *deleted;
+	struct tidhash_hash *deleting;
 	BufferAccessStrategy bas;
 	HnswNeighborTuple ntup;
 	HnswElementData highestPoint;

src/hnswbuild.c

@@ -719,7 +719,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	/* Get support functions */
 	HnswInitSupport(&buildstate->support, index);
 
-	InitGraph(&buildstate->graphData, NULL, (Size) maintenance_work_mem * 1024L);
+	InitGraph(&buildstate->graphData, NULL, maintenance_work_mem * (Size) 1024);
 	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
@@ -956,7 +956,7 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	/* Leave space for other objects in shared memory */
 	/* Docker has a default limit of 64 MB for shm_size */
 	/* which happens to be the default value of maintenance_work_mem */
-	esthnswarea = maintenance_work_mem * 1024L;
+	esthnswarea = maintenance_work_mem * (Size) 1024;
 	estother = 3 * 1024 * 1024;
 	if (esthnswarea > estother)
 		esthnswarea -= estother;

src/hnswvacuum.c

@@ -19,12 +19,12 @@
 #endif
 
 /*
- * Check if deleted list contains an index TID
+ * Check if deletion list contains an element
  */
 static bool
-DeletedContains(tidhash_hash * deleted, ItemPointer indextid)
+DeletingElement(tidhash_hash * deleting, ItemPointer indextid)
 {
-	return tidhash_lookup(deleted, *indextid) != NULL;
+	return tidhash_lookup(deleting, *indextid) != NULL;
 }
 
 /*
@@ -80,6 +80,14 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 			if (!HnswIsElementTuple(etup))
 				continue;
 
+			/*
+			 * Skip deleted tuples. It is important they are not added to the
+			 * deletion list to avoid false positives in NeedsUpdated and
+			 * ConfirmRepaired.
+			 */
+			if (etup->deleted)
+				continue;
+
 			if (ItemPointerIsValid(&etup->heaptids[0]))
 			{
 				for (int i = 0; i < HNSW_HEAPTIDS; i++)
@@ -113,13 +121,13 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 			if (!ItemPointerIsValid(&etup->heaptids[0]))
 			{
-				ItemPointerData ip;
+				ItemPointerData indextid;
 				bool		found;
 
-				/* Add to deleted list */
-				ItemPointerSet(&ip, blkno, offno);
+				/* Add to deletion list */
+				ItemPointerSet(&indextid, blkno, offno);
 
-				tidhash_insert(vacuumstate->deleted, ip, &found);
+				tidhash_insert(vacuumstate->deleting, indextid, &found);
 				Assert(!found);
 			}
 			else if (etup->level > highestLevel)
@@ -192,8 +200,8 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
 		if (!ItemPointerIsValid(indextid))
 			continue;
 
-		/* Check if in deleted list */
-		if (DeletedContains(vacuumstate->deleted, indextid))
+		/* Check if in deletion list */
+		if (DeletingElement(vacuumstate->deleting, indextid))
 		{
 			needsUpdated = true;
 			break;
@@ -202,13 +210,9 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
 
 	/* Also update if layer 0 is not full */
 	/* This could indicate too many candidates being deleted during insert */
-	if (!needsUpdated)
-	{
-		/* Keep clang-tidy happy */
-		Assert(ntup->count > 0);
-
+	/* There should always be more than zero indextids, but check for safety */
+	if (!needsUpdated && ntup->count > 0)
 		needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
-	}
 
 	UnlockReleaseBuffer(buf);
 
@@ -331,7 +335,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 
 		ItemPointerSet(&epData, entryPoint->blkno, entryPoint->offno);
 
-		if (DeletedContains(vacuumstate->deleted, &epData))
+		if (DeletingElement(vacuumstate->deleting, &epData))
 		{
 			/*
 			 * Replace the entry point with the highest point. If highest
@@ -417,6 +421,10 @@ RepairGraph(HnswVacuumState * vacuumstate)
 			if (!HnswIsElementTuple(etup))
 				continue;
 
+			/* Skip deleted tuples */
+			if (etup->deleted)
+				continue;
+
 			/* Skip updating neighbors if being deleted */
 			if (!ItemPointerIsValid(&etup->heaptids[0]))
 				continue;
@@ -531,6 +539,10 @@ ConfirmRepaired(HnswVacuumState * vacuumstate)
 			if (!HnswIsElementTuple(etup))
 				continue;
 
+			/* Skip deleted tuples */
+			if (etup->deleted)
+				continue;
+
 			/* Skip if being deleted */
 			if (!ItemPointerIsValid(&etup->heaptids[0]))
 				continue;
@@ -561,8 +573,8 @@ ConfirmRepaired(HnswVacuumState * vacuumstate)
 				if (!ItemPointerIsValid(indextid))
 					continue;
 
-				/* Check if in deleted list */
-				if (DeletedContains(vacuumstate->deleted, indextid))
+				/* Check if in deletion list */
+				if (DeletingElement(vacuumstate->deleting, indextid))
 					elog(ERROR, "hnsw graph not repaired");
 			}
 
@@ -588,10 +600,15 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	BufferAccessStrategy bas = vacuumstate->bas;
 
 	/*
-	 * Wait for index scans to complete. Scans before this point may contain
-	 * tuples about to be deleted. Scans after this point will not, since the
-	 * graph has been repaired.
+	 * Wait for inserts and index scans to complete. Inserts and scans before
+	 * this point may visit tuples about to be deleted. Inserts and scans
+	 * after this point will not, since the graph has been repaired.
 	 */
+	LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+	UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
+
+	ConfirmRepaired(vacuumstate);
+
 	LockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
 	UnlockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
 
@@ -739,7 +756,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 	HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
 
 	/* Create hash table */
-	vacuumstate->deleted = tidhash_create(CurrentMemoryContext, 256, NULL);
+	vacuumstate->deleting = tidhash_create(CurrentMemoryContext, 256, NULL);
 }
 
 /*
@@ -748,7 +765,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
 static void
 FreeVacuumState(HnswVacuumState * vacuumstate)
 {
-	tidhash_destroy(vacuumstate->deleted);
+	tidhash_destroy(vacuumstate->deleting);
 	FreeAccessStrategy(vacuumstate->bas);
 	pfree(vacuumstate->ntup);
 	MemoryContextDelete(vacuumstate->tmpCtx);
@@ -771,10 +788,7 @@ hnswbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	/* Pass 2: Repair graph */
 	HnswBench("RepairGraph", RepairGraph(&vacuumstate));
 
-	/* Pass 3: Confirm repaired */
-	HnswBench("ConfirmRepaired", ConfirmRepaired(&vacuumstate));
-
-	/* Pass 4: Mark as deleted */
+	/* Passes 3 and 4: Confirm repaired and mark as deleted */
 	HnswBench("MarkDeleted", MarkDeleted(&vacuumstate));
 
 	FreeVacuumState(&vacuumstate);

src/ivfbuild.c

@@ -152,18 +152,7 @@ SampleRows(IvfflatBuildState * buildstate)
 
 	/* Normalize if needed */
 	if (buildstate->kmeansnormprocinfo != NULL)
-	{
-		VectorArray samples = buildstate->samples;
-
-		for (int i = 0; i < samples->length; i++)
-		{
-			Datum		value = PointerGetDatum(VectorArrayGet(samples, i));
-			Datum		normValue = IvfflatNormValue(buildstate->typeInfo, buildstate->collation, value);
-
-			VectorArraySet(samples, i, DatumGetPointer(normValue));
-			pfree(DatumGetPointer(normValue));
-		}
-	}
+		IvfflatNormVectors(buildstate->typeInfo, buildstate->collation, buildstate->samples, buildstate->tmpCtx);
 }
 
 /*
@@ -399,8 +388,14 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 
 	buildstate->slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsVirtual);
 
-	/* TODO Ensure within maintenance_work_mem */
-	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions, buildstate->typeInfo->itemSize(buildstate->dimensions));
+	buildstate->memoryUsed = 0;
+	buildstate->itemsize = buildstate->typeInfo->itemSize(buildstate->dimensions);
+
+	buildstate->memoryUsed += VECTOR_ARRAY_SIZE(buildstate->lists, buildstate->itemsize);
+	IvfflatCheckMemoryUsage(buildstate->memoryUsed);
+	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions, buildstate->itemsize);
+
+	/* TODO Move allocation to page creation */
 	buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
 
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
@@ -443,19 +438,27 @@ ComputeCenters(IvfflatBuildState * buildstate)
 
 	pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
 
+	/* Skip samples for unlogged table */
+	if (buildstate->heap == NULL)
+		numSamples = 1;
+	else
+	{
+		int64		maxTuples = (int64) RelationGetNumberOfBlocks(buildstate->heap) * MaxHeapTuplesPerPage;
+
 		/* Target 50 samples per list, with at least 10000 samples */
 		/* The number of samples has a large effect on index build time */
 		numSamples = buildstate->lists * 50;
 		if (numSamples < 10000)
 			numSamples = 10000;
 
-	/* Skip samples for unlogged table */
-	if (buildstate->heap == NULL)
-		numSamples = 1;
+		/* Save memory since will not have more than max tuples */
+		numSamples = Max(Min(numSamples, maxTuples), 1);
+	}
 
 	/* Sample rows */
-	/* TODO Ensure within maintenance_work_mem */
-	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions, buildstate->centers->itemsize);
+	buildstate->memoryUsed += VECTOR_ARRAY_SIZE(numSamples, buildstate->itemsize);
+	IvfflatCheckMemoryUsage(buildstate->memoryUsed);
+	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions, buildstate->itemsize);
 	if (buildstate->heap != NULL)
 	{
 		IvfflatBench("sample rows", SampleRows(buildstate));
@@ -470,7 +473,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	}
 
 	/* Calculate centers */
-	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers, buildstate->typeInfo));
+	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers, buildstate->typeInfo, buildstate->memoryUsed));
 
 	/* Free samples before we allocate more memory */
 	VectorArrayFree(buildstate->samples);

src/ivfflat.h

@@ -204,6 +204,7 @@ typedef struct IvfflatBuildState
 	VectorArray samples;
 	VectorArray centers;
 	ListInfo   *listInfo;
+	Size		itemsize;
 
 #ifdef IVFFLAT_KMEANS_DEBUG
 	double		inertia;
@@ -223,6 +224,7 @@ typedef struct IvfflatBuildState
 	TupleTableSlot *slot;
 
 	/* Memory */
+	Size		memoryUsed;
 	MemoryContext tmpCtx;
 
 	/* Parallel builds */
@@ -303,22 +305,32 @@ typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 static inline Pointer
 VectorArrayGet(VectorArray arr, int offset)
 {
+	if (offset >= arr->maxlen)
+		elog(ERROR, "safety check failed");
+
 	return ((char *) arr->items) + (offset * arr->itemsize);
 }
 
 static inline void
 VectorArraySet(VectorArray arr, int offset, Pointer val)
 {
-	memcpy(VectorArrayGet(arr, offset), val, VARSIZE_ANY(val));
+	Size		size = VARSIZE_ANY(val);
+
+	if (size > arr->itemsize)
+		elog(ERROR, "safety check failed");
+
+	memcpy(VectorArrayGet(arr, offset), val, size);
 }
 
 /* Methods */
 VectorArray VectorArrayInit(int maxlen, int dimensions, Size itemsize);
 void		VectorArrayFree(VectorArray arr);
-void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo);
+void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo, Size memoryUsed);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 Datum		IvfflatNormValue(const IvfflatTypeInfo * typeInfo, Oid collation, Datum value);
 bool		IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
+void		IvfflatNormVectors(const IvfflatTypeInfo * typeInfo, Oid collation, VectorArray arr, MemoryContext tmpCtx);
+void		IvfflatCheckMemoryUsage(Size totalSize);
 int			IvfflatGetLists(Relation index);
 void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

src/ivfkmeans.c

@@ -8,6 +8,7 @@
 #include "fmgr.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 #include "utils/relcache.h"
 
@@ -99,36 +100,50 @@ NormCenters(const IvfflatTypeInfo * typeInfo, Oid collation, VectorArray centers
 	MemoryContext normCtx = AllocSetContextCreate(CurrentMemoryContext,
 												  "Ivfflat norm temporary context",
 												  ALLOCSET_DEFAULT_SIZES);
-	MemoryContext oldCtx = MemoryContextSwitchTo(normCtx);
 
-	for (int j = 0; j < centers->length; j++)
-	{
-		Datum		center = PointerGetDatum(VectorArrayGet(centers, j));
-		Datum		newCenter = IvfflatNormValue(typeInfo, collation, center);
-		Size		size = VARSIZE_ANY(DatumGetPointer(newCenter));
-
-		if (size > centers->itemsize)
-			elog(ERROR, "safety check failed");
-
-		memcpy(DatumGetPointer(center), DatumGetPointer(newCenter), size);
-		MemoryContextReset(normCtx);
-	}
-
-	MemoryContextSwitchTo(oldCtx);
+	IvfflatNormVectors(typeInfo, collation, centers, normCtx);
 	MemoryContextDelete(normCtx);
 }
 
 /*
- * Quick approach if we have no data
+ * Check if vector array contains a vector
+ */
+static bool
+VectorArrayContains(VectorArray arr, Pointer val)
+{
+	Datum		d = PointerGetDatum(val);
+
+	for (int i = 0; i < arr->length; i++)
+	{
+		if (datumIsEqual(d, PointerGetDatum(VectorArrayGet(arr, i)), false, -1))
+			return true;
+	}
+	return false;
+}
+
+/*
+ * Quick approach if we have little data
  */
 static void
-RandomCenters(Relation index, VectorArray centers, const IvfflatTypeInfo * typeInfo)
+QuickCenters(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo)
 {
 	int			dimensions = centers->dim;
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 	Oid			collation = index->rd_indcollation[0];
 	float	   *x = (float *) palloc(sizeof(float) * dimensions);
 
+	/* Fill with unique samples (already normalized) */
+	for (int i = 0; i < samples->length; i++)
+	{
+		Pointer		sample = VectorArrayGet(samples, i);
+
+		if (!VectorArrayContains(centers, sample))
+		{
+			VectorArraySet(centers, centers->length, sample);
+			centers->length++;
+		}
+	}
+
 	/* Fill with random data */
 	while (centers->length < centers->maxlen)
 	{
@@ -258,7 +273,7 @@ ComputeNewCenters(VectorArray samples, float *agg, VectorArray newCenters, int *
  * https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf
  */
 static void
-ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo)
+ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo, Size memoryUsed)
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
@@ -277,8 +292,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const Ivff
 	float	   *newcdist;
 
 	/* Calculate allocation sizes */
-	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->itemsize);
-	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->itemsize);
 	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, centers->itemsize);
 	Size		aggSize = sizeof(float) * (int64) numCenters * dimensions;
 	Size		centerCountsSize = sizeof(int) * numCenters;
@@ -290,18 +303,13 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const Ivff
 	Size		newcdistSize = sizeof(float) * numCenters;
 
 	/* Calculate total size */
-	Size		totalSize = samplesSize + centersSize + newCentersSize + aggSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
+	Size		totalSize = memoryUsed + newCentersSize + aggSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
 
 	/* Check memory requirements */
-	/* Add one to error message to ceil */
-	if (totalSize > (Size) maintenance_work_mem * 1024L)
-		ereport(ERROR,
-				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
-				 errmsg("memory required is %zu MB, maintenance_work_mem is %d MB",
-						totalSize / (1024 * 1024) + 1, maintenance_work_mem / 1024)));
+	IvfflatCheckMemoryUsage(totalSize);
 
 	/* Ensure indexing does not overflow */
-	if (numCenters * numCenters > INT_MAX)
+	if (numCenters > INT_MAX / numCenters)
 		elog(ERROR, "Indexing overflow detected. Please report a bug.");
 
 	/* Set support functions */
@@ -562,17 +570,17 @@ CheckCenters(Relation index, VectorArray centers, const IvfflatTypeInfo * typeIn
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo)
+IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo, Size memoryUsed)
 {
 	MemoryContext kmeansCtx = AllocSetContextCreate(CurrentMemoryContext,
 													"Ivfflat kmeans temporary context",
 													ALLOCSET_DEFAULT_SIZES);
 	MemoryContext oldCtx = MemoryContextSwitchTo(kmeansCtx);
 
-	if (samples->length == 0)
-		RandomCenters(index, centers, typeInfo);
+	if (samples->length <= centers->maxlen)
+		QuickCenters(index, samples, centers, typeInfo);
 	else
-		ElkanKmeans(index, samples, centers, typeInfo);
+		ElkanKmeans(index, samples, centers, typeInfo, memoryUsed);
 
 	CheckCenters(index, centers, typeInfo);
 

src/ivfutils.c

@@ -6,7 +6,9 @@
 #include "halfutils.h"
 #include "halfvec.h"
 #include "ivfflat.h"
+#include "miscadmin.h"
 #include "storage/bufmgr.h"
+#include "utils/memutils.h"
 #include "utils/relcache.h"
 #include "utils/varbit.h"
 #include "vector.h"
@@ -21,11 +23,15 @@
 VectorArray
 VectorArrayInit(int maxlen, int dimensions, Size itemsize)
 {
-	VectorArray res = palloc(sizeof(VectorArrayData));
+	VectorArray res;
+
+	if (maxlen < 1 || dimensions < 1 || itemsize == 0)
+		elog(ERROR, "cannot create vector array");
 
 	/* Ensure items are aligned to prevent UB */
 	itemsize = MAXALIGN(itemsize);
 
+	res = palloc(sizeof(VectorArrayData));
 	res->length = 0;
 	res->maxlen = maxlen;
 	res->dim = dimensions;
@@ -88,6 +94,40 @@ IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
 	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
 }
 
+/*
+ * Normalize vectors
+ */
+void
+IvfflatNormVectors(const IvfflatTypeInfo * typeInfo, Oid collation, VectorArray arr, MemoryContext tmpCtx)
+{
+	MemoryContext oldCtx = MemoryContextSwitchTo(tmpCtx);
+
+	for (int i = 0; i < arr->length; i++)
+	{
+		Datum		value = PointerGetDatum(VectorArrayGet(arr, i));
+		Datum		newValue = IvfflatNormValue(typeInfo, collation, value);
+
+		VectorArraySet(arr, i, DatumGetPointer(newValue));
+		MemoryContextReset(tmpCtx);
+	}
+
+	MemoryContextSwitchTo(oldCtx);
+}
+
+/*
+ * Check memory usage
+ */
+void
+IvfflatCheckMemoryUsage(Size totalSize)
+{
+	/* Add one to error message to ceil */
+	if (totalSize > maintenance_work_mem * (Size) 1024)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("memory required is %zu MB, maintenance_work_mem is %d MB",
+						totalSize / (1024 * 1024) + 1, maintenance_work_mem / 1024)));
+}
+
 /*
  * New buffer
  */

src/vector.c

@@ -40,7 +40,7 @@
 #endif
 
 #if PG_VERSION_NUM >= 180000
-PG_MODULE_MAGIC_EXT(.name = "vector", .version = "0.8.3");
+PG_MODULE_MAGIC_EXT(.name = "vector", .version = "0.8.5");
 #else
 PG_MODULE_MAGIC;
 #endif

test/expected/hnsw_bit.out

@@ -49,3 +49,11 @@ CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
 CREATE INDEX ON t USING hnsw ((val::bit(64001)) bit_hamming_ops);
 ERROR:  column cannot have more than 64000 dimensions for hnsw index
 DROP TABLE t;
+-- dimensions
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+DROP TABLE t;
+CREATE TABLE t (val bit(64001));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+ERROR:  column cannot have more than 64000 dimensions for hnsw index
+DROP TABLE t;

test/expected/hnsw_halfvec.out

@@ -100,3 +100,11 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::halfvec)) t
 (1 row)
 
 DROP TABLE t;
+-- dimensions
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING hnsw (val halfvec_l2_ops);
+DROP TABLE t;
+CREATE TABLE t (val halfvec(4001));
+CREATE INDEX ON t USING hnsw (val halfvec_l2_ops);
+ERROR:  column cannot have more than 4000 dimensions for hnsw index
+DROP TABLE t;

test/expected/hnsw_vector.out

@@ -161,6 +161,7 @@ ERROR:  value 1001 out of bounds for option "ef_construction"
 DETAIL:  Valid values are between "4" and "1000".
 CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
 ERROR:  ef_construction must be greater than or equal to 2 * m
+DROP TABLE t;
 SHOW hnsw.ef_search;
  hnsw.ef_search 
 ----------------
@@ -198,4 +199,11 @@ SET hnsw.scan_mem_multiplier = 0;
 ERROR:  0 is outside the valid range for parameter "hnsw.scan_mem_multiplier" (1 .. 1000)
 SET hnsw.scan_mem_multiplier = 1001;
 ERROR:  1001 is outside the valid range for parameter "hnsw.scan_mem_multiplier" (1 .. 1000)
+-- dimensions
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+DROP TABLE t;
+CREATE TABLE t (val vector(2001));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+ERROR:  column cannot have more than 2000 dimensions for hnsw index
 DROP TABLE t;

test/expected/ivfflat_bit.out

@@ -35,3 +35,32 @@ NOTICE:  ivfflat index created with little data
 DETAIL:  This will cause low recall.
 HINT:  Drop the index until the table has more data.
 DROP TABLE t;
+-- dimensions
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+CREATE TABLE t (val bit(64001));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+ERROR:  column cannot have more than 64000 dimensions for ivfflat index
+DROP TABLE t;
+-- memory
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;
+SET maintenance_work_mem = '29MB';
+CREATE TABLE t (val bit(64000));
+INSERT INTO t (val) VALUES (B'0'::bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;

test/expected/ivfflat_halfvec.out

@@ -82,3 +82,32 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::halfvec)) t
 (1 row)
 
 DROP TABLE t;
+-- dimensions
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+CREATE TABLE t (val halfvec(4001));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+ERROR:  column cannot have more than 4000 dimensions for ivfflat index
+DROP TABLE t;
+-- memory
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;
+SET maintenance_work_mem = '6MB';
+CREATE TABLE t (val halfvec(4000));
+INSERT INTO t (val) VALUES (array_fill(0, ARRAY[4000]));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;

test/expected/ivfflat_vector.out

@@ -143,6 +143,7 @@ DETAIL:  Valid values are between "1" and "32768".
 CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
 ERROR:  value 32769 out of bounds for option "lists"
 DETAIL:  Valid values are between "1" and "32768".
+DROP TABLE t;
 SHOW ivfflat.probes;
  ivfflat.probes 
 ----------------
@@ -172,4 +173,32 @@ SET ivfflat.max_probes = 0;
 ERROR:  0 is outside the valid range for parameter "ivfflat.max_probes" (1 .. 32768)
 SET ivfflat.max_probes = 32769;
 ERROR:  32769 is outside the valid range for parameter "ivfflat.max_probes" (1 .. 32768)
+-- dimensions
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
 DROP TABLE t;
+CREATE TABLE t (val vector(2001));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+ERROR:  column cannot have more than 2000 dimensions for ivfflat index
+DROP TABLE t;
+-- memory
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;
+SET maintenance_work_mem = '5MB';
+CREATE TABLE t (val vector(2000));
+INSERT INTO t (val) VALUES (array_fill(0, ARRAY[2000]));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;
+RESET maintenance_work_mem;

test/sql/hnsw_bit.sql

@@ -33,3 +33,13 @@ CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
 CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
 CREATE INDEX ON t USING hnsw ((val::bit(64001)) bit_hamming_ops);
 DROP TABLE t;
+
+-- dimensions
+
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+DROP TABLE t;
+
+CREATE TABLE t (val bit(64001));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+DROP TABLE t;

test/sql/hnsw_halfvec.sql

@@ -56,3 +56,13 @@ SELECT * FROM t ORDER BY val <+> '[3,3,3]';
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::halfvec)) t2;
 
 DROP TABLE t;
+
+-- dimensions
+
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING hnsw (val halfvec_l2_ops);
+DROP TABLE t;
+
+CREATE TABLE t (val halfvec(4001));
+CREATE INDEX ON t USING hnsw (val halfvec_l2_ops);
+DROP TABLE t;

test/sql/hnsw_vector.sql

@@ -95,23 +95,28 @@ CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
 CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
 CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
 CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
+DROP TABLE t;
 
 SHOW hnsw.ef_search;
-
 SET hnsw.ef_search = 0;
 SET hnsw.ef_search = 1001;
 
 SHOW hnsw.iterative_scan;
-
 SET hnsw.iterative_scan = on;
 
 SHOW hnsw.max_scan_tuples;
-
 SET hnsw.max_scan_tuples = 0;
 
 SHOW hnsw.scan_mem_multiplier;
-
 SET hnsw.scan_mem_multiplier = 0;
 SET hnsw.scan_mem_multiplier = 1001;
 
+-- dimensions
+
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
+DROP TABLE t;
+
+CREATE TABLE t (val vector(2001));
+CREATE INDEX ON t USING hnsw (val vector_l2_ops);
 DROP TABLE t;

test/sql/ivfflat_bit.sql

@@ -21,3 +21,28 @@ CREATE INDEX ON t USING ivfflat ((val::bit(3)) bit_hamming_ops) WITH (lists = 1)
 CREATE INDEX ON t USING ivfflat ((val::bit(64001)) bit_hamming_ops) WITH (lists = 1);
 CREATE INDEX ON t USING ivfflat ((val::bit(2)) bit_hamming_ops) WITH (lists = 5);
 DROP TABLE t;
+
+-- dimensions
+
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+DROP TABLE t;
+
+CREATE TABLE t (val bit(64001));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+DROP TABLE t;
+
+-- memory
+
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;
+
+SET maintenance_work_mem = '29MB';
+CREATE TABLE t (val bit(64000));
+INSERT INTO t (val) VALUES (B'0'::bit(64000));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;

test/sql/ivfflat_halfvec.sql

@@ -43,3 +43,28 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::halfvec)) t2;
 
 DROP TABLE t;
+
+-- dimensions
+
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+DROP TABLE t;
+
+CREATE TABLE t (val halfvec(4001));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+DROP TABLE t;
+
+-- memory
+
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val halfvec(4000));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;
+
+SET maintenance_work_mem = '6MB';
+CREATE TABLE t (val halfvec(4000));
+INSERT INTO t (val) VALUES (array_fill(0, ARRAY[4000]));
+CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;

test/sql/ivfflat_vector.sql

@@ -81,19 +81,40 @@ DROP TABLE t;
 CREATE TABLE t (val vector(3));
 CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
 CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
+DROP TABLE t;
 
 SHOW ivfflat.probes;
-
 SET ivfflat.probes = 0;
 SET ivfflat.probes = 32769;
 
 SHOW ivfflat.iterative_scan;
-
 SET ivfflat.iterative_scan = on;
 
 SHOW ivfflat.max_probes;
-
 SET ivfflat.max_probes = 0;
 SET ivfflat.max_probes = 32769;
 
+-- dimensions
+
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
 DROP TABLE t;
+
+CREATE TABLE t (val vector(2001));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+DROP TABLE t;
+
+-- memory
+
+SET maintenance_work_mem = '1MB';
+CREATE TABLE t (val vector(2000));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;
+
+SET maintenance_work_mem = '5MB';
+CREATE TABLE t (val vector(2000));
+INSERT INTO t (val) VALUES (array_fill(0, ARRAY[2000]));
+CREATE INDEX ON t USING ivfflat (val vector_l2_ops);
+DROP TABLE t;
+RESET maintenance_work_mem;

test/t/046_hnsw_vacuum_scan.pl

@@ -0,0 +1,38 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector($dim));");
+$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
+$node->safe_psql("postgres",
+    "INSERT INTO tst (v) SELECT ARRAY[$array_sql] FROM generate_series(1, 1000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
+$node->safe_psql("postgres", "DELETE FROM tst");
+
+# Test HNSW_SCAN_LOCK at the beginning of MarkDeleted is effective
+$node->pgbench(
+    "--no-vacuum --client=5 --transactions=1000",
+    0,
+    [qr{actually processed}],
+    [qr{^$}],
+    "concurrent SELECTs and VACUUM",
+    {
+        "046_hnsw_vacuum_scan_select\@1000" => "SELECT i FROM tst ORDER BY v <-> '[0,0,0]' LIMIT 10;",
+        "046_hnsw_vacuum_scan_vacuum\@1" => "VACUUM tst;"
+    }
+);
+
+done_testing();

test/t/047_hnsw_vacuum_insert.pl

@@ -0,0 +1,39 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector($dim));");
+$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
+$node->safe_psql("postgres",
+    "INSERT INTO tst (v) SELECT ARRAY[$array_sql] FROM generate_series(1, 1000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops);");
+
+# Test no "hnsw graph not repaired" errors
+$node->pgbench(
+    "--no-vacuum --client=5 --transactions=1000",
+    0,
+    [qr{actually processed}],
+    [qr{^$}],
+    "concurrent INSERTs, DELETEs, SELECTs, and VACUUM",
+    {
+        "047_hnsw_vacuum_insert_insert\@500" => "INSERT INTO tst (v) VALUES (ARRAY[$array_sql]);",
+        "047_hnsw_vacuum_insert_delete\@500" => "DELETE FROM tst WHERE i = (SELECT i FROM tst LIMIT 1);",
+        "047_hnsw_vacuum_insert_select\@20" => "SELECT i FROM tst ORDER BY v <-> (SELECT ARRAY[$array_sql]::vector) LIMIT 10;",
+        "047_hnsw_vacuum_insert_vacuum\@1" => "VACUUM tst;"
+    }
+);
+
+done_testing();

test/t/048_ivfflat_vacuum_insert.pl

@@ -0,0 +1,39 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+my $dim = 3;
+my $array_sql = join(",", ('random()') x $dim);
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table and index
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v vector($dim));");
+$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
+$node->safe_psql("postgres",
+    "INSERT INTO tst (v) SELECT ARRAY[$array_sql] FROM generate_series(1, 1000) i;"
+);
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10);");
+
+# Test no errors
+$node->pgbench(
+    "--no-vacuum --client=5 --transactions=1500",
+    0,
+    [qr{actually processed}],
+    [qr{^$}],
+    "concurrent INSERTs, DELETEs, SELECTs, and VACUUM",
+    {
+        "048_ivfflat_vacuum_insert_insert\@500" => "INSERT INTO tst (v) VALUES (ARRAY[$array_sql]);",
+        "048_ivfflat_vacuum_insert_delete\@500" => "DELETE FROM tst WHERE i = (SELECT i FROM tst LIMIT 1);",
+        "048_ivfflat_vacuum_insert_select\@500" => "SELECT i FROM tst ORDER BY v <-> (SELECT ARRAY[$array_sql]::vector) LIMIT 10;",
+        "048_ivfflat_vacuum_insert_vacuum\@1" => "VACUUM tst;"
+    }
+);
+
+done_testing();

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.8.3'
+default_version = '0.8.5'
 module_pathname = '$libdir/vector'
 relocatable = true