Improved naming [skip ci]

Co-authored-by: Bhagyesh Chaturvedi <bhagyeshc@google.com>

.github/workflows/build.yml

@@ -97,7 +97,7 @@ jobs:
         with:
           postgres-version: ${{ matrix.postgres }}
       - run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat" && ^
+          call "C:\Program Files\Microsoft Visual Studio\${{ matrix.os == 'windows-2025' && 18 || 2022 }}\Enterprise\VC\Auxiliary\Build\vcvars64.bat" && ^
           nmake /NOLOGO /F Makefile.win && ^
           nmake /NOLOGO /F Makefile.win install && ^
           nmake /NOLOGO /F Makefile.win installcheck ${{ matrix.postgres != 17 && 'PG_REGRESS=$(PGROOT)\bin\pg_regress' || '' }} && ^

CHANGELOG.md

@@ -1,3 +1,13 @@
+## 0.8.4 (unreleased)
+
+- Fixed `hnsw graph not repaired` error with HNSW vacuuming
+- Fixed possible error with inserts during HNSW vacuuming
+
+## 0.8.3 (2026-06-17)
+
+- Fixed possible index corruption with HNSW vacuuming
+- Fixed performance regression with Hamming distance and Jaccard distance with Postgres 18
+
 ## 0.8.2 (2026-02-25)
 
 - Fixed buffer overflow with parallel HNSW index build - [more info](https://github.com/pgvector/pgvector/issues/959)

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#v0.8.2 /tmp/pgvector
+ADD https://github.com/pgvector/pgvector.git#v0.8.3 /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": "0.8.2",
+	"version": "0.8.3",
 	"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": "0.8.2",
+			"version": "0.8.3",
 			"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 = 0.8.2
+EXTVERSION = 0.8.3
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.2
+EXTVERSION = 0.8.3
 
 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 v0.8.2 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.3 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 v0.8.2 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -1151,23 +1151,23 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 
 Supported tags are:
 
-- `pg18-trixie`, `0.8.2-pg18-trixie`
-- `pg18-bookworm`, `0.8.2-pg18-bookworm`, `pg18`, `0.8.2-pg18`
-- `pg17-trixie`, `0.8.2-pg17-trixie`
-- `pg17-bookworm`, `0.8.2-pg17-bookworm`, `pg17`, `0.8.2-pg17`
-- `pg16-trixie`, `0.8.2-pg16-trixie`
-- `pg16-bookworm`, `0.8.2-pg16-bookworm`, `pg16`, `0.8.2-pg16`
-- `pg15-trixie`, `0.8.2-pg15-trixie`
-- `pg15-bookworm`, `0.8.2-pg15-bookworm`, `pg15`, `0.8.2-pg15`
-- `pg14-trixie`, `0.8.2-pg14-trixie`
-- `pg14-bookworm`, `0.8.2-pg14-bookworm`, `pg14`, `0.8.2-pg14`
-- `pg13-trixie`, `0.8.2-pg13-trixie`
-- `pg13-bookworm`, `0.8.2-pg13-bookworm`, `pg13`, `0.8.2-pg13`
+- `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`
 
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.8.2 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --pull --build-arg PG_MAJOR=18 -t myuser/pgvector .
 ```
@@ -1329,7 +1329,7 @@ make clean && PG_CFLAGS="-DUSE_ASSERT_CHECKING" make && make install
 To enable benchmarking:
 
 ```sh
-make clean && PG_CFLAGS="-DIVFFLAT_BENCH" make && make install
+make clean && PG_CFLAGS="-DHNSW_BENCH -DIVFFLAT_BENCH" make && make install
 ```
 
 To show memory usage:

sql/vector--0.8.2--0.8.3.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.8.3'" to load this file. \quit

src/bitutils.c

@@ -31,10 +31,12 @@
 #define BIT_TARGET_CLONES
 #endif
 
-/* Use built-ins when possible for inlining */
-#if defined(HAVE__BUILTIN_POPCOUNT) && defined(HAVE_LONG_INT_64)
+/* Use built-ins when possible for Postgres < 19 for inlining */
+#if PG_VERSION_NUM >= 190000
+#define popcount64(x) pg_popcount64(x)
+#elif defined(HAVE__BUILTIN_POPCOUNT) && (defined(HAVE_LONG_INT_64) || SIZEOF_LONG == 8)
 #define popcount64(x) __builtin_popcountl(x)
-#elif defined(HAVE__BUILTIN_POPCOUNT) && defined(HAVE_LONG_LONG_INT_64)
+#elif defined(HAVE__BUILTIN_POPCOUNT) && (defined(HAVE_LONG_LONG_INT_64) || SIZEOF_LONG_LONG == 8)
 #define popcount64(x) __builtin_popcountll(x)
 #elif !defined(_MSC_VER)
 /* Fails to resolve with MSVC */

src/hnsw.h

@@ -18,6 +18,10 @@
 #include "utils/sampling.h"
 #include "vector.h"
 
+#ifdef HNSW_BENCH
+#include "portability/instr_time.h"
+#endif
+
 #if PG_VERSION_NUM >= 190000
 typedef Pointer Item;
 #endif
@@ -78,6 +82,21 @@ typedef Pointer Item;
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
 #define HnswPageGetMeta(page)	((HnswMetaPageData *) PageGetContents(page))
 
+#ifdef HNSW_BENCH
+#define HnswBench(name, code) \
+	do { \
+		instr_time	start; \
+		instr_time	duration; \
+		INSTR_TIME_SET_CURRENT(start); \
+		(code); \
+		INSTR_TIME_SET_CURRENT(duration); \
+		INSTR_TIME_SUBTRACT(duration, start); \
+		elog(INFO, "%s: %.3f ms", name, INSTR_TIME_GET_MILLISEC(duration)); \
+	} while (0)
+#else
+#define HnswBench(name, code) (code)
+#endif
+
 #if PG_VERSION_NUM >= 150000
 #define RandomDouble() pg_prng_double(&pg_global_prng_state)
 #define SeedRandom(seed) pg_prng_seed(&pg_global_prng_state, seed)
@@ -408,10 +427,11 @@ typedef struct HnswVacuumState
 	HnswSupport support;
 
 	/* Variables */
-	struct tidhash_hash *deleted;
+	struct tidhash_hash *deleting;
 	BufferAccessStrategy bas;
 	HnswNeighborTuple ntup;
 	HnswElementData highestPoint;
+	HnswElementData fallbackPoint;
 
 	/* Memory */
 	MemoryContext tmpCtx;

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/hnswutils.c

@@ -546,6 +546,9 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, Hns
 
 	Assert(HnswIsElementTuple(etup));
 
+	if (unlikely(etup->deleted))
+		elog(ERROR, "cannot load deleted element");
+
 	/* Calculate distance */
 	if (distance != NULL)
 	{

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;
 }
 
 /*
@@ -37,17 +37,20 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 {
 	BlockNumber blkno = HNSW_HEAD_BLKNO;
 	HnswElement highestPoint = &vacuumstate->highestPoint;
+	HnswElement fallbackPoint = &vacuumstate->fallbackPoint;
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
-	HnswElement entryPoint = HnswGetEntryPoint(vacuumstate->index);
 	IndexBulkDeleteResult *stats = vacuumstate->stats;
 
-	/* Store separately since highestPoint.level is uint8 */
+	/* Store separately since HnswElement level is uint8 */
 	int			highestLevel = -1;
+	int			fallbackLevel = -1;
 
-	/* Initialize highest point */
+	/* Initialize highest point and fallback point */
 	highestPoint->blkno = InvalidBlockNumber;
 	highestPoint->offno = InvalidOffsetNumber;
+	fallbackPoint->blkno = InvalidBlockNumber;
+	fallbackPoint->offno = InvalidOffsetNumber;
 
 	while (BlockNumberIsValid(blkno))
 	{
@@ -77,6 +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++)
@@ -110,23 +121,40 @@ 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 && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
+			else if (etup->level > highestLevel)
 			{
-				/* Keep track of highest non-entry point */
+				if (BlockNumberIsValid(highestPoint->blkno))
+				{
+					/* Current highest point becomes fallback */
+					fallbackPoint->blkno = highestPoint->blkno;
+					fallbackPoint->offno = highestPoint->offno;
+					fallbackPoint->level = highestPoint->level;
+					fallbackLevel = highestLevel;
+				}
+
+				/* Keep track of highest point */
 				highestPoint->blkno = blkno;
 				highestPoint->offno = offno;
 				highestPoint->level = etup->level;
 				highestLevel = etup->level;
 			}
+			else if (etup->level > fallbackLevel)
+			{
+				/* Keep track of second highest point */
+				fallbackPoint->blkno = blkno;
+				fallbackPoint->offno = offno;
+				fallbackPoint->level = etup->level;
+				fallbackLevel = etup->level;
+			}
 		}
 
 		blkno = HnswPageGetOpaque(page)->nextblkno;
@@ -138,6 +166,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 		UnlockReleaseBuffer(buf);
 	}
+
+#ifdef HNSW_MEMORY
+	elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(CurrentMemoryContext, true) / 1024);
+#endif
 }
 
 /*
@@ -168,8 +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;
@@ -178,13 +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);
 
@@ -269,12 +297,27 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		/* Get a shared lock */
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
 
-		/* Load element */
-		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
+		/* Get latest entry point */
+		entryPoint = HnswGetEntryPoint(index);
 
-		/* Repair if needed */
-		if (NeedsUpdated(vacuumstate, highestPoint))
-			RepairGraphElement(vacuumstate, highestPoint, HnswGetEntryPoint(index));
+		/* Use fallback point if highest point is entry point */
+		if (entryPoint != NULL && entryPoint->blkno == highestPoint->blkno && entryPoint->offno == highestPoint->offno)
+		{
+			highestPoint = &vacuumstate->fallbackPoint;
+
+			if (!BlockNumberIsValid(highestPoint->blkno))
+				highestPoint = NULL;
+		}
+
+		if (highestPoint != NULL)
+		{
+			/* Load element */
+			HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
+
+			/* Repair if needed */
+			if (NeedsUpdated(vacuumstate, highestPoint))
+				RepairGraphElement(vacuumstate, highestPoint, entryPoint);
+		}
 
 		/* Release lock */
 		UnlockPage(index, HNSW_UPDATE_LOCK, ShareLock);
@@ -292,7 +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
@@ -378,6 +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;
@@ -441,6 +488,103 @@ RepairGraph(HnswVacuumState * vacuumstate)
 		/* Reset memory context */
 		MemoryContextSwitchTo(oldCtx);
 		MemoryContextReset(vacuumstate->tmpCtx);
+
+#ifdef HNSW_VACUUM_PROGRESS
+		if (!BlockNumberIsValid(blkno) || (blkno - HNSW_HEAD_BLKNO) % 1000 == 0)
+		{
+			BlockNumber totalBlocks = RelationGetNumberOfBlocks(index);
+			BlockNumber currentBlocks = BlockNumberIsValid(blkno) ? blkno : totalBlocks;
+
+			elog(INFO, "hnsw vacuum progress: %.1f%%", 100.0 * currentBlocks / totalBlocks);
+		}
+#endif
+	}
+}
+
+/*
+ * Confirm graph was repaired
+ */
+static void
+ConfirmRepaired(HnswVacuumState * vacuumstate)
+{
+	BlockNumber blkno = HNSW_HEAD_BLKNO;
+	Relation	index = vacuumstate->index;
+	BufferAccessStrategy bas = vacuumstate->bas;
+
+	while (BlockNumberIsValid(blkno))
+	{
+		Buffer		buf;
+		Page		page;
+		OffsetNumber offno;
+		OffsetNumber maxoffno;
+
+		vacuum_delay_point();
+
+		buf = ReadBufferExtended(index, MAIN_FORKNUM, blkno, RBM_NORMAL, bas);
+		LockBuffer(buf, BUFFER_LOCK_SHARE);
+		page = BufferGetPage(buf);
+		maxoffno = PageGetMaxOffsetNumber(page);
+
+		/* Iterate over nodes */
+		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
+		{
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			HnswNeighborTuple ntup;
+			Buffer		nbuf;
+			Page		npage;
+			BlockNumber neighborPage;
+			OffsetNumber neighborOffno;
+
+			/* Skip neighbor tuples */
+			if (!HnswIsElementTuple(etup))
+				continue;
+
+			/* Skip deleted tuples */
+			if (etup->deleted)
+				continue;
+
+			/* Skip if being deleted */
+			if (!ItemPointerIsValid(&etup->heaptids[0]))
+				continue;
+
+			/* Get neighbor page */
+			neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
+			neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
+
+			if (neighborPage == blkno)
+			{
+				nbuf = buf;
+				npage = page;
+			}
+			else
+			{
+				nbuf = ReadBufferExtended(index, MAIN_FORKNUM, neighborPage, RBM_NORMAL, bas);
+				LockBuffer(nbuf, BUFFER_LOCK_SHARE);
+				npage = BufferGetPage(nbuf);
+			}
+
+			ntup = (HnswNeighborTuple) PageGetItem(npage, PageGetItemId(npage, neighborOffno));
+
+			/* Check neighbors */
+			for (int i = 0; i < ntup->count; i++)
+			{
+				ItemPointer indextid = &ntup->indextids[i];
+
+				if (!ItemPointerIsValid(indextid))
+					continue;
+
+				/* Check if in deletion list */
+				if (DeletingElement(vacuumstate->deleting, indextid))
+					elog(ERROR, "hnsw graph not repaired");
+			}
+
+			if (nbuf != buf)
+				UnlockReleaseBuffer(nbuf);
+		}
+
+		blkno = HnswPageGetOpaque(page)->nextblkno;
+
+		UnlockReleaseBuffer(buf);
 	}
 }
 
@@ -456,10 +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);
 
@@ -607,7 +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);
 }
 
 /*
@@ -616,7 +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);
@@ -634,13 +783,13 @@ hnswbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	InitVacuumState(&vacuumstate, info, stats, callback, callback_state);
 
 	/* Pass 1: Remove heap TIDs */
-	RemoveHeapTids(&vacuumstate);
+	HnswBench("RemoveHeapTids", RemoveHeapTids(&vacuumstate));
 
 	/* Pass 2: Repair graph */
-	RepairGraph(&vacuumstate);
+	HnswBench("RepairGraph", RepairGraph(&vacuumstate));
 
-	/* Pass 3: Mark as deleted */
-	MarkDeleted(&vacuumstate);
+	/* 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);
 }
 
 /*
@@ -399,7 +388,14 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 
 	buildstate->slot = MakeSingleTupleTableSlot(buildstate->sortdesc, &TTSOpsVirtual);
 
-	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,
@@ -453,8 +449,9 @@ ComputeCenters(IvfflatBuildState * buildstate)
 		numSamples = 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));
@@ -469,7 +466,7 @@ 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

@@ -99,22 +99,8 @@ 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);
 }
 
@@ -258,7 +244,7 @@ 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 +263,6 @@ 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 +274,13 @@ 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 */
@@ -562,7 +541,7 @@ 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",
@@ -572,7 +551,7 @@ IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const Iv
 	if (samples->length == 0)
 		RandomCenters(index, 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"
@@ -88,6 +90,40 @@ 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 = "0.8.2");
+PG_MODULE_MAGIC_EXT(.name = "vector", .version = "0.8.3");
 #else
 PG_MODULE_MAGIC;
 #endif

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.8.2'
+default_version = '0.8.3'
 module_pathname = '$libdir/vector'
 relocatable = true