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 }}
@@ -92,12 +94,12 @@ 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 }}
       - 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' || '' }} && ^
@@ -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 +1,18 @@
+## 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
+- 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#v0.8.5 /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": "0.8.2",
+	"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": "0.8.2",
+			"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 = 0.8.2
+EXTVERSION = 0.8.5
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.2
+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

@@ -11,6 +11,8 @@ Store your vectors with the rest of your data. Supports:
 
 Plus [ACID](https://en.wikipedia.org/wiki/ACID) compliance, point-in-time recovery, JOINs, and all of the other [great features](https://www.postgresql.org/about/) of Postgres
 
+Have a lot of vectors? Use [quantization](#scaling) to scale
+
 [![Build Status](https://github.com/pgvector/pgvector/actions/workflows/build.yml/badge.svg)](https://github.com/pgvector/pgvector/actions)
 
 ## Installation
@@ -21,7 +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 v0.8.5 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -38,7 +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 v0.8.5 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -314,6 +316,8 @@ For a large number of workers, you may need to increase `max_parallel_workers` (
 
 The [index options](#index-options) also have a significant impact on build time (use the defaults unless seeing low recall)
 
+Use [binary quantization](#binary-quantization) for faster build times at scale
+
 ### Indexing Progress
 
 Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING)
@@ -443,13 +447,7 @@ Exact indexes work well for conditions that match a low percentage of rows. Othe
 CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
 ```
 
-With approximate indexes, filtering is applied *after* the index is scanned. If a condition matches 10% of rows, with HNSW and the default `hnsw.ef_search` of 40, only 4 rows will match on average. For more rows, increase `hnsw.ef_search`.
+With approximate indexes, filtering is applied *after* the index is scanned. If a condition matches 10% of rows, with HNSW and the default `hnsw.ef_search` of 40, only 4 rows will match on average. For more rows, enable [iterative index scans](#iterative-index-scans), which will automatically scan more of the index when needed.
-
-```sql
-SET hnsw.ef_search = 200;
-```
-
-Starting with 0.8.0, you can enable [iterative index scans](#iterative-index-scans), which will automatically scan more of the index when needed.
 
 ```sql
 SET hnsw.iterative_scan = strict_order;
@@ -467,6 +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`).
@@ -673,6 +681,10 @@ SHOW shared_buffers;
 
 Be sure to restart Postgres for changes to take effect.
 
+### Storing
+
+Use the `halfvec` type instead of `vector` for a smaller working set.
+
 ### Loading
 
 Use `COPY` for bulk loading data ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/loading/example.py)).
@@ -687,6 +699,8 @@ Add any indexes *after* loading the initial data for best performance.
 
 See index build time for [HNSW](#index-build-time) and [IVFFlat](#index-build-time-1).
 
+Use [binary quantization](#binary-quantization) for smaller indexes and faster build times at scale.
+
 In production environments, create indexes concurrently to avoid blocking writes.
 
 ```sql
@@ -717,6 +731,8 @@ SELECT * FROM items ORDER BY embedding <#> '[3,1,2]' LIMIT 5;
 
 #### Approximate Search
 
+Use [binary quantization](#binary-quantization) with re-ranking to keep indexes in-memory at scale.
+
 To speed up queries with an IVFFlat index, increase the number of inverted lists (at the expense of recall).
 
 ```sql
@@ -732,21 +748,20 @@ REINDEX INDEX CONCURRENTLY index_name;
 VACUUM table_name;
 ```
 
+## Scaling
+
+For a smaller working set:
+
+1. Use the `halfvec` type instead of `vector` for tables
+2. Use [binary quantization](#binary-quantization) for indexes (with re-ranking for search)
+
+Scale vertically by increasing memory, CPU, and storage on a single instance. Use existing tools to [tune parameters](#tuning) and [monitor performance](#monitoring).
+
+Scale horizontally with [replicas](https://www.postgresql.org/docs/current/hot-standby.html), or use [Citus](https://github.com/citusdata/citus), [PgDog](https://github.com/pgdogdev/pgdog), or another approach for sharding ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/citus/example.py)).
+
 ## Monitoring
 
-Monitor performance with [pg_stat_statements](https://www.postgresql.org/docs/current/pgstatstatements.html) (be sure to add it to `shared_preload_libraries`).
+Use existing tools like [pg_stat_statements](https://www.postgresql.org/docs/current/pgstatstatements.html) or [PgHero](https://github.com/ankane/pghero) to monitor performance.
-
-```sql
-CREATE EXTENSION pg_stat_statements;
-```
-
-Get the most time-consuming queries with:
-
-```sql
-SELECT query, calls, ROUND((total_plan_time + total_exec_time) / calls) AS avg_time_ms,
-    ROUND((total_plan_time + total_exec_time) / 60000) AS total_time_min
-    FROM pg_stat_statements ORDER BY total_plan_time + total_exec_time DESC LIMIT 20;
-```
 
 Monitor recall by comparing results from approximate search with exact search.
 
@@ -757,14 +772,6 @@ SELECT ...
 COMMIT;
 ```
 
-## Scaling
-
-Scale pgvector the same way you scale Postgres.
-
-Scale vertically by increasing memory, CPU, and storage on a single instance. Use existing tools to [tune parameters](#tuning) and [monitor performance](#monitoring).
-
-Scale horizontally with [replicas](https://www.postgresql.org/docs/current/hot-standby.html), or use [Citus](https://github.com/citusdata/citus) or another approach for sharding ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/citus/example.py)).
-
 ## Languages
 
 Use pgvector from any language with a Postgres client. You can even generate and store vectors in one language and query them in another.
@@ -878,6 +885,8 @@ No, but like other index types, you’ll likely see better performance if they d
 SELECT pg_size_pretty(pg_relation_size('index_name'));
 ```
 
+Use [half-precision indexing](#half-precision-indexing) or [binary quantization](#binary-quantization) for smaller indexes.
+
 ## Troubleshooting
 
 #### Why isn’t a query using an index?
@@ -1152,23 +1161,23 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 
 Supported tags are:
 
-- `pg18-trixie`, `0.8.2-pg18-trixie`
+- `pg18-trixie`, `0.8.5-pg18-trixie`
-- `pg18-bookworm`, `0.8.2-pg18-bookworm`, `pg18`, `0.8.2-pg18`
+- `pg18-bookworm`, `0.8.5-pg18-bookworm`, `pg18`, `0.8.5-pg18`
-- `pg17-trixie`, `0.8.2-pg17-trixie`
+- `pg17-trixie`, `0.8.5-pg17-trixie`
-- `pg17-bookworm`, `0.8.2-pg17-bookworm`, `pg17`, `0.8.2-pg17`
+- `pg17-bookworm`, `0.8.5-pg17-bookworm`, `pg17`, `0.8.5-pg17`
-- `pg16-trixie`, `0.8.2-pg16-trixie`
+- `pg16-trixie`, `0.8.5-pg16-trixie`
-- `pg16-bookworm`, `0.8.2-pg16-bookworm`, `pg16`, `0.8.2-pg16`
+- `pg16-bookworm`, `0.8.5-pg16-bookworm`, `pg16`, `0.8.5-pg16`
-- `pg15-trixie`, `0.8.2-pg15-trixie`
+- `pg15-trixie`, `0.8.5-pg15-trixie`
-- `pg15-bookworm`, `0.8.2-pg15-bookworm`, `pg15`, `0.8.2-pg15`
+- `pg15-bookworm`, `0.8.5-pg15-bookworm`, `pg15`, `0.8.5-pg15`
-- `pg14-trixie`, `0.8.2-pg14-trixie`
+- `pg14-trixie`, `0.8.5-pg14-trixie`
-- `pg14-bookworm`, `0.8.2-pg14-bookworm`, `pg14`, `0.8.2-pg14`
+- `pg14-bookworm`, `0.8.5-pg14-bookworm`, `pg14`, `0.8.5-pg14`
-- `pg13-trixie`, `0.8.2-pg13-trixie`
+- `pg13-trixie`, `0.8.5-pg13-trixie`
-- `pg13-bookworm`, `0.8.2-pg13-bookworm`, `pg13`, `0.8.2-pg13`
+- `pg13-bookworm`, `0.8.5-pg13-bookworm`, `pg13`, `0.8.5-pg13`
 
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.8.2 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 .
 ```
@@ -1330,7 +1339,7 @@ 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

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/bitutils.c

@@ -31,10 +31,12 @@
 #define BIT_TARGET_CLONES
 #endif
 
-/* Use built-ins when possible for inlining */
+/* Use built-ins when possible for Postgres < 19 for inlining */
-#if defined(HAVE__BUILTIN_POPCOUNT) && defined(HAVE_LONG_INT_64)
+#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)
@@ -372,13 +391,6 @@ typedef union
 	ItemPointerData indextid;
 }			HnswUnvisited;
 
-typedef struct HnswReadStreamData
-{
-	HnswUnvisited *unvisited;
-	int			unvisitedLength;
-	int			visited;
-}			HnswReadStreamData;
-
 typedef struct HnswScanOpaqueData
 {
 	const		HnswTypeInfo *typeInfo;
@@ -415,10 +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;
@@ -434,13 +447,13 @@ bool		HnswCheckNorm(HnswSupport * support, Datum value);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);
-List	   *HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples, bool maintenance);
+List	   *HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples);
 HnswElement HnswGetEntryPoint(Relation index);
 void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
 void	   *HnswAlloc(HnswAllocator * allocator, Size size);
 HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
 HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
-void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing, bool maintenance);
+void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing);
 HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement entryPoint, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec);
 void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
 void		HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);

src/hnswbuild.c

@@ -470,7 +470,7 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 	}
 
 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false, true);
+	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false);
 
 	/* Update graph in memory */
 	UpdateGraphInMemory(support, element, m, entryPoint, buildstate);
@@ -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/hnswinsert.c

@@ -731,7 +731,7 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPo
 	}
 
 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false, building);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false);
 
 	/* Update graph on disk */
 	UpdateGraphOnDisk(index, support, element, m, entryPoint, building);

src/hnswscan.c

@@ -48,11 +48,11 @@ GetScanItems(IndexScanDesc scan, Datum value)
 
 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, NULL, NULL, true, NULL, false);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
-	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, support, m, false, NULL, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &so->discarded : NULL, true, &so->tuples, false);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, support, m, false, NULL, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &so->discarded : NULL, true, &so->tuples);
 }
 
 /*
@@ -83,7 +83,7 @@ ResumeScanItems(IndexScanDesc scan)
 		ep = lappend(ep, sc);
 	}
 
-	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples, false);
+	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
 }
 
 /*

src/hnswutils.c

@@ -21,10 +21,6 @@
 #include "varatt.h"
 #endif
 
-#if PG_VERSION_NUM >= 190000
-#include "storage/read_stream.h"
-#endif
-
 #if PG_VERSION_NUM < 170000
 static inline uint64
 murmurhash64(uint64 data)
@@ -535,12 +531,14 @@ HnswGetDistance(Datum a, Datum b, HnswSupport * support)
  * Load an element and optionally get its distance from q
  */
 static void
-HnswLoadElementImpl(Buffer buf, OffsetNumber offno, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance, HnswElement * element)
+HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance, HnswElement * element)
 {
+	Buffer		buf;
 	Page		page;
 	HnswElementTuple etup;
 
 	/* Read vector */
+	buf = ReadBuffer(index, blkno);
 	LockBuffer(buf, BUFFER_LOCK_SHARE);
 	page = BufferGetPage(buf);
 
@@ -548,6 +546,9 @@ HnswLoadElementImpl(Buffer buf, OffsetNumber offno, double *distance, HnswQuery
 
 	Assert(HnswIsElementTuple(etup));
 
+	if (unlikely(etup->deleted))
+		elog(ERROR, "cannot load deleted element");
+
 	/* Calculate distance */
 	if (distance != NULL)
 	{
@@ -561,7 +562,7 @@ HnswLoadElementImpl(Buffer buf, OffsetNumber offno, double *distance, HnswQuery
 	if (distance == NULL || maxDistance == NULL || *distance < *maxDistance)
 	{
 		if (*element == NULL)
-			*element = HnswInitElementFromBlock(BufferGetBlockNumber(buf), offno);
+			*element = HnswInitElementFromBlock(blkno, offno);
 
 		HnswLoadElementFromTuple(*element, etup, true, loadVec);
 	}
@@ -575,9 +576,7 @@ HnswLoadElementImpl(Buffer buf, OffsetNumber offno, double *distance, HnswQuery
 void
 HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
 {
-	Buffer		buf = ReadBuffer(index, element->blkno);
+	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
-
-	HnswLoadElementImpl(buf, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
 }
 
 /*
@@ -817,31 +816,11 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
 	}
 }
 
-#if PG_VERSION_NUM >= 190000
-/*
- * Get next block number for read stream
- */
-static BlockNumber
-HnswReadStreamNextBlock(ReadStream *stream, void *callback_private_data, void *per_buffer_data)
-{
-	HnswReadStreamData *streamData = callback_private_data;
-	OffsetNumber *offno = per_buffer_data;
-	HnswUnvisited *uv;
-
-	if (streamData->visited == streamData->unvisitedLength)
-		return InvalidBlockNumber;
-
-	uv = &streamData->unvisited[streamData->visited++];
-	*offno = ItemPointerGetOffsetNumber(&uv->indextid);
-	return ItemPointerGetBlockNumber(&uv->indextid);
-}
-#endif
-
 /*
  * Algorithm 2 from paper
  */
 List *
-HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples, bool maintenance)
+HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation index, HnswSupport * support, int m, bool inserting, HnswElement skipElement, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
 {
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
@@ -856,21 +835,6 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 	int			unvisitedLength;
 	bool		inMemory = index == NULL;
 
-#if PG_VERSION_NUM >= 190000
-	HnswReadStreamData streamData;
-	ReadStream *stream = NULL;
-
-	if (!inMemory)
-	{
-		int			flags = READ_STREAM_USE_BATCHING;
-
-		if (maintenance)
-			flags |= READ_STREAM_MAINTENANCE;
-
-		stream = read_stream_begin_relation(flags, NULL, index, MAIN_FORKNUM, HnswReadStreamNextBlock, &streamData, sizeof(OffsetNumber));
-	}
-#endif
-
 	if (v == NULL)
 	{
 		v = &vh;
@@ -933,23 +897,13 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 		if (inMemory)
 			HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, v, lc, localNeighborhood, neighborhoodSize);
 		else
-		{
 			HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, v, index, m, lm, lc);
 
-#if PG_VERSION_NUM >= 190000
-			read_stream_resume(stream);
-
-			streamData.unvisited = unvisited;
-			streamData.unvisitedLength = unvisitedLength;
-			streamData.visited = 0;
-#endif
-		}
-
 		/* OK to count elements instead of tuples */
 		if (tuples != NULL)
 			(*tuples) += unvisitedLength;
 
-		for (int i = 0;; i++)
+		for (int i = 0; i < unvisitedLength; i++)
 		{
 			HnswElement eElement;
 			HnswSearchCandidate *e;
@@ -960,40 +914,18 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 
 			if (inMemory)
 			{
-				if (i == unvisitedLength)
-					break;
-
 				eElement = unvisited[i].element;
 				eDistance = GetElementDistance(base, eElement, q, support);
 			}
 			else
 			{
-				Buffer		buf;
+				ItemPointer indextid = &unvisited[i].indextid;
-				OffsetNumber offno;
+				BlockNumber blkno = ItemPointerGetBlockNumber(indextid);
-
+				OffsetNumber offno = ItemPointerGetOffsetNumber(indextid);
-#if PG_VERSION_NUM >= 190000
-				void	   *offnoPtr;
-
-				buf = read_stream_next_buffer(stream, &offnoPtr);
-
-				if (!BufferIsValid(buf))
-					break;
-
-				offno = *((OffsetNumber *) offnoPtr);
-#else
-				ItemPointer indextid;
-
-				if (i == unvisitedLength)
-					break;
-
-				indextid = &unvisited[i].indextid;
-				buf = ReadBuffer(index, ItemPointerGetBlockNumber(indextid));
-				offno = ItemPointerGetOffsetNumber(indextid);
-#endif
 
 				/* Avoid any allocations if not adding */
 				eElement = NULL;
-				HnswLoadElementImpl(buf, offno, &eDistance, q, index, support, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);
+				HnswLoadElementImpl(blkno, offno, &eDistance, q, index, support, inserting, alwaysAdd || discarded != NULL ? NULL : &f->distance, &eElement);
 
 				if (eElement == NULL)
 					continue;
@@ -1049,11 +981,6 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 		w = lappend(w, sc);
 	}
 
-#if PG_VERSION_NUM >= 190000
-	if (!inMemory)
-		read_stream_end(stream);
-#endif
-
 	return w;
 }
 
@@ -1349,7 +1276,7 @@ PrecomputeHash(char *base, HnswElement element)
  * Algorithm 1 from paper
  */
 void
-HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing, bool maintenance)
+HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing)
 {
 	List	   *ep;
 	List	   *w;
@@ -1376,7 +1303,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL, maintenance);
+		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
@@ -1395,7 +1322,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		List	   *lw = NIL;
 		ListCell   *lc2;
 
-		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL, maintenance);
+		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
 
 		/* Convert search candidates to candidates */
 		foreach(lc2, w)

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 */
+				/* Add to deletion list */
-				ItemPointerSet(&ip, blkno, offno);
+				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 */
+		/* Check if in deletion list */
-		if (DeletedContains(vacuumstate->deleted, indextid))
+		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)
+	/* There should always be more than zero indextids, but check for safety */
-	{
+	if (!needsUpdated && ntup->count > 0)
-		/* Keep clang-tidy happy */
-		Assert(ntup->count > 0);
-
 		needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
-	}
 
 	UnlockReleaseBuffer(buf);
 
@@ -218,7 +246,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	element->heaptidsLength = 0;
 
 	/* Find neighbors for element, skipping itself */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, true, true);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, true);
 
 	/* Zero memory for each element */
 	MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
@@ -269,12 +297,27 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		/* Get a shared lock */
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
 
-		/* Load element */
+		/* Get latest entry point */
-		HnswLoadElement(highestPoint, NULL, NULL, index, support, true, NULL);
+		entryPoint = HnswGetEntryPoint(index);
 
-		/* Repair if needed */
+		/* Use fallback point if highest point is entry point */
-		if (NeedsUpdated(vacuumstate, highestPoint))
+		if (entryPoint != NULL && entryPoint->blkno == highestPoint->blkno && entryPoint->offno == highestPoint->offno)
-			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);
@@ -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
+	 * Wait for inserts and index scans to complete. Inserts and scans before
-	 * tuples about to be deleted. Scans after this point will not, since the
+	 * this point may visit tuples about to be deleted. Inserts and scans
-	 * graph has been repaired.
+	 * 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 */
+	/* Passes 3 and 4: Confirm repaired and mark as deleted */
-	MarkDeleted(&vacuumstate);
+	HnswBench("MarkDeleted", MarkDeleted(&vacuumstate));
 
 	FreeVacuumState(&vacuumstate);
 

src/ivfbuild.c

@@ -145,24 +145,14 @@ SampleRows(IvfflatBuildState * buildstate)
 	{
 		BlockNumber targblock = BlockSampler_Next(&buildstate->bs);
 
+		/* Set anyvisible to false like table_index_build_scan */
 		table_index_build_range_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-									 false, true, false, targblock, 1, SampleCallback, (void *) buildstate, NULL);
+									 false, false, false, targblock, 1, SampleCallback, (void *) buildstate, NULL);
 	}
 
 	/* Normalize if needed */
 	if (buildstate->kmeansnormprocinfo != NULL)
-	{
+		IvfflatNormVectors(buildstate->typeInfo, buildstate->collation, buildstate->samples, buildstate->tmpCtx);
-		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));
-		}
-	}
 }
 
 /*
@@ -398,7 +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,
@@ -441,19 +438,27 @@ ComputeCenters(IvfflatBuildState * buildstate)
 
 	pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
 
-	/* 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;
+	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;
+
+		/* 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->memoryUsed += VECTOR_ARRAY_SIZE(numSamples, buildstate->itemsize);
-	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions, buildstate->centers->itemsize);
+	IvfflatCheckMemoryUsage(buildstate->memoryUsed);
+	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions, buildstate->itemsize);
 	if (buildstate->heap != NULL)
 	{
 		IvfflatBench("sample rows", SampleRows(buildstate));
@@ -468,7 +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++)
+	IvfflatNormVectors(typeInfo, collation, centers, normCtx);
-	{
-		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);
 	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 */
+	IvfflatCheckMemoryUsage(totalSize);
-	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)));
 
 	/* 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)
+	if (samples->length <= centers->maxlen)
-		RandomCenters(index, centers, typeInfo);
+		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/sparsevec.c

@@ -182,10 +182,10 @@ sparsevec_isspace(char ch)
 static int
 CompareIndices(const void *a, const void *b)
 {
-	if (((SparseInputElement *) a)->index < ((SparseInputElement *) b)->index)
+	if (((const SparseInputElement *) a)->index < ((const SparseInputElement *) b)->index)
 		return -1;
 
-	if (((SparseInputElement *) a)->index > ((SparseInputElement *) b)->index)
+	if (((const SparseInputElement *) a)->index > ((const SparseInputElement *) b)->index)
 		return 1;
 
 	return 0;

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 = "0.8.5");
 #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 = '0.8.2'
+default_version = '0.8.5'
 module_pathname = '$libdir/vector'
 relocatable = true