Added benchmarking for HNSW vacuuming [skip ci]

CHANGELOG.md

@@ -1,6 +1,5 @@
-## 0.8.3 (2026-06-17)
+## 0.8.3 (unreleased)
 
-- 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)

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#v0.8.2 /tmp/pgvector
 
 RUN apt-get update && \
 		apt-mark hold locales && \

META.json

@@ -2,7 +2,7 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.8.3",
+	"version": "0.8.2",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.8.3",
+			"version": "0.8.2",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.3
+EXTVERSION = 0.8.2
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.3
+EXTVERSION = 0.8.2
 
 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 v0.8.2 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 v0.8.2 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`, `0.8.3-pg18-trixie`
+- `pg18-trixie`, `0.8.2-pg18-trixie`
-- `pg18-bookworm`, `0.8.3-pg18-bookworm`, `pg18`, `0.8.3-pg18`
+- `pg18-bookworm`, `0.8.2-pg18-bookworm`, `pg18`, `0.8.2-pg18`
-- `pg17-trixie`, `0.8.3-pg17-trixie`
+- `pg17-trixie`, `0.8.2-pg17-trixie`
-- `pg17-bookworm`, `0.8.3-pg17-bookworm`, `pg17`, `0.8.3-pg17`
+- `pg17-bookworm`, `0.8.2-pg17-bookworm`, `pg17`, `0.8.2-pg17`
-- `pg16-trixie`, `0.8.3-pg16-trixie`
+- `pg16-trixie`, `0.8.2-pg16-trixie`
-- `pg16-bookworm`, `0.8.3-pg16-bookworm`, `pg16`, `0.8.3-pg16`
+- `pg16-bookworm`, `0.8.2-pg16-bookworm`, `pg16`, `0.8.2-pg16`
-- `pg15-trixie`, `0.8.3-pg15-trixie`
+- `pg15-trixie`, `0.8.2-pg15-trixie`
-- `pg15-bookworm`, `0.8.3-pg15-bookworm`, `pg15`, `0.8.3-pg15`
+- `pg15-bookworm`, `0.8.2-pg15-bookworm`, `pg15`, `0.8.2-pg15`
-- `pg14-trixie`, `0.8.3-pg14-trixie`
+- `pg14-trixie`, `0.8.2-pg14-trixie`
-- `pg14-bookworm`, `0.8.3-pg14-bookworm`, `pg14`, `0.8.3-pg14`
+- `pg14-bookworm`, `0.8.2-pg14-bookworm`, `pg14`, `0.8.2-pg14`
-- `pg13-trixie`, `0.8.3-pg13-trixie`
+- `pg13-trixie`, `0.8.2-pg13-trixie`
-- `pg13-bookworm`, `0.8.3-pg13-bookworm`, `pg13`, `0.8.3-pg13`
+- `pg13-bookworm`, `0.8.2-pg13-bookworm`, `pg13`, `0.8.2-pg13`
 
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.8.3 https://github.com/pgvector/pgvector.git
+git clone --branch v0.8.2 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --pull --build-arg PG_MAJOR=18 -t myuser/pgvector .
 ```

sql/vector--0.8.2--0.8.3.sql

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

src/hnsw.h

@@ -431,7 +431,6 @@ typedef struct HnswVacuumState
 	BufferAccessStrategy bas;
 	HnswNeighborTuple ntup;
 	HnswElementData highestPoint;
-	HnswElementData fallbackPoint;
 
 	/* Memory */
 	MemoryContext tmpCtx;

src/hnswutils.c

@@ -546,9 +546,6 @@ 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

@@ -37,20 +37,17 @@ 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 HnswElement level is uint8 */
+	/* Store separately since highestPoint.level is uint8 */
 	int			highestLevel = -1;
-	int			fallbackLevel = -1;
 
-	/* Initialize highest point and fallback point */
+	/* Initialize highest point */
 	highestPoint->blkno = InvalidBlockNumber;
 	highestPoint->offno = InvalidOffsetNumber;
-	fallbackPoint->blkno = InvalidBlockNumber;
-	fallbackPoint->offno = InvalidOffsetNumber;
 
 	while (BlockNumberIsValid(blkno))
 	{
@@ -122,31 +119,14 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 				tidhash_insert(vacuumstate->deleted, ip, &found);
 				Assert(!found);
 			}
-			else if (etup->level > highestLevel)
+			else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
 			{
-				if (BlockNumberIsValid(highestPoint->blkno))
+				/* Keep track of highest non-entry point */
-				{
-					/* 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;
@@ -158,10 +138,6 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 		UnlockReleaseBuffer(buf);
 	}
-
-#ifdef HNSW_MEMORY
-	elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(CurrentMemoryContext, true) / 1024);
-#endif
 }
 
 /*
@@ -293,27 +269,12 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		/* Get a shared lock */
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
 
-		/* Get latest entry point */
+		/* Load element */
-		entryPoint = HnswGetEntryPoint(index);
+		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
 
-		/* Use fallback point if highest point is entry point */
+		/* Repair if needed */
-		if (entryPoint != NULL && entryPoint->blkno == highestPoint->blkno && entryPoint->offno == highestPoint->offno)
+		if (NeedsUpdated(vacuumstate, highestPoint))
-		{
+			RepairGraphElement(vacuumstate, highestPoint, HnswGetEntryPoint(index));
-			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);
@@ -480,99 +441,6 @@ 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 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 deleted list */
-				if (DeletedContains(vacuumstate->deleted, indextid))
-					elog(ERROR, "hnsw graph not repaired");
-			}
-
-			if (nbuf != buf)
-				UnlockReleaseBuffer(nbuf);
-		}
-
-		blkno = HnswPageGetOpaque(page)->nextblkno;
-
-		UnlockReleaseBuffer(buf);
 	}
 }
 
@@ -771,10 +639,7 @@ hnswbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
 	/* Pass 2: Repair graph */
 	HnswBench("RepairGraph", RepairGraph(&vacuumstate));
 
-	/* Pass 3: Confirm repaired */
+	/* Pass 3: Mark as deleted */
-	HnswBench("ConfirmRepaired", ConfirmRepaired(&vacuumstate));
-
-	/* Pass 4: Mark as deleted */
 	HnswBench("MarkDeleted", MarkDeleted(&vacuumstate));
 
 	FreeVacuumState(&vacuumstate);

src/ivfbuild.c

@@ -399,7 +399,6 @@ 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->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
 

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 = "0.8.2");
 #else
 PG_MODULE_MAGIC;
 #endif

vector.control

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