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@v5
+      - 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@v5
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -72,7 +74,7 @@ jobs:
           tar xf $TAG.tar.gz
           mv postgres-$TAG postgres
         env:
-          TAG: ${{ matrix.postgres == 18 && 'REL_18_0' || 'REL_14_19' }}
+          TAG: ${{ matrix.postgres == 18 && 'REL_18_2' || 'REL_14_21' }}
       - run: make prove_installcheck PROVE_FLAGS="-I ./postgres/src/test/perl -I ./test/perl"
         env:
           PERL5LIB: /Users/runner/perl5/lib/perl5
@@ -92,12 +94,12 @@ jobs:
           - postgres: 14
             os: windows-2022
     steps:
-      - uses: actions/checkout@v5
+      - 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@v5
+      - uses: actions/checkout@v7
       - uses: ankane/setup-postgres-valgrind@v1
         with:
           postgres-version: 18

CHANGELOG.md

@@ -1,9 +1,21 @@
-## 0.9.0 (unreleased)
+## 0.8.5 (2026-07-08)
 
-- Added support for inline filtering with HNSW
+- Reduced memory usage for small tables for IVFFlat index builds
 
-## 0.8.2 (unreleased)
+## 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)
 - Improved `install` target on Windows
 - Fixed `Index Searches` in `EXPLAIN` output for Postgres 18
 

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#v0.8.1 /tmp/pgvector
+ADD https://github.com/pgvector/pgvector.git#v0.8.5 /tmp/pgvector
 
 RUN apt-get update && \
 		apt-mark hold locales && \

LICENSE

@@ -1,4 +1,4 @@
-Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
+Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
 
 Portions Copyright (c) 1994, The Regents of the University of California
 

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": "0.8.1",
+	"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": "0.8.1",
+			"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 = 0.8.1
+EXTVERSION = 0.8.5
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*--*.sql)
@@ -27,6 +27,11 @@ ifneq ($(filter ppc64%, $(shell uname -m)), )
 	OPTFLAGS =
 endif
 
+# RISC-V64 doesn't support -march=native
+ifeq ($(shell uname -m), riscv64)
+	OPTFLAGS =
+endif
+
 # For auto-vectorization:
 # - GCC (needs -ftree-vectorize OR -O3) - https://gcc.gnu.org/projects/tree-ssa/vectorization.html
 # - Clang (could use pragma instead) - https://llvm.org/docs/Vectorizers.html

Makefile.win

@@ -1,5 +1,5 @@
 EXTENSION = vector
-EXTVERSION = 0.8.1
+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 v0.8.1 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 v0.8.1 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`.
-
-```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.
+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.iterative_scan = strict_order;
@@ -467,10 +465,14 @@ 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);
 ```
 
-Or a composite HNSW index (added in 0.9.0)
+## 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 INDEX ON items USING hnsw (embedding vector_l2_ops, category_id);
+CREATE TABLE items (customer_id int, embedding vector(3)) PARTITION BY LIST(customer_id);
 ```
 
 ## Iterative Index Scans
@@ -679,6 +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)).
@@ -693,6 +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
@@ -723,6 +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
@@ -738,21 +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`).
-
-```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;
-```
+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.
 
 Monitor recall by comparing results from approximate search with exact search.
 
@@ -763,14 +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.
@@ -782,6 +783,7 @@ Algol | [pgvector-algol](https://github.com/pgvector/pgvector-algol)
 C | [pgvector-c](https://github.com/pgvector/pgvector-c)
 C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
 C#, F#, Visual Basic | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
+COBOL | [pgvector-cobol](https://github.com/pgvector/pgvector-cobol)
 Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
 D | [pgvector-d](https://github.com/pgvector/pgvector-d)
 Dart | [pgvector-dart](https://github.com/pgvector/pgvector-dart)
@@ -809,6 +811,7 @@ Raku | [pgvector-raku](https://github.com/pgvector/pgvector-raku)
 Ruby | [pgvector-ruby](https://github.com/pgvector/pgvector-ruby), [Neighbor](https://github.com/ankane/neighbor)
 Rust | [pgvector-rust](https://github.com/pgvector/pgvector-rust)
 Swift | [pgvector-swift](https://github.com/pgvector/pgvector-swift)
+Tcl | [pgvector-tcl](https://github.com/pgvector/pgvector-tcl)
 Zig | [pgvector-zig](https://github.com/pgvector/pgvector-zig)
 
 ## Frequently Asked Questions
@@ -882,6 +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?
@@ -1156,23 +1161,23 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 
 Supported tags are:
 
-- `pg18-trixie`, `0.8.1-pg18-trixie`
-- `pg18-bookworm`, `0.8.1-pg18-bookworm`, `pg18`, `0.8.1-pg18`
-- `pg17-trixie`, `0.8.1-pg17-trixie`
-- `pg17-bookworm`, `0.8.1-pg17-bookworm`, `pg17`, `0.8.1-pg17`
-- `pg16-trixie`, `0.8.1-pg16-trixie`
-- `pg16-bookworm`, `0.8.1-pg16-bookworm`, `pg16`, `0.8.1-pg16`
-- `pg15-trixie`, `0.8.1-pg15-trixie`
-- `pg15-bookworm`, `0.8.1-pg15-bookworm`, `pg15`, `0.8.1-pg15`
-- `pg14-trixie`, `0.8.1-pg14-trixie`
-- `pg14-bookworm`, `0.8.1-pg14-bookworm`, `pg14`, `0.8.1-pg14`
-- `pg13-trixie`, `0.8.1-pg13-trixie`
-- `pg13-bookworm`, `0.8.1-pg13-bookworm`, `pg13`, `0.8.1-pg13`
+- `pg18-trixie`, `0.8.5-pg18-trixie`
+- `pg18-bookworm`, `0.8.5-pg18-bookworm`, `pg18`, `0.8.5-pg18`
+- `pg17-trixie`, `0.8.5-pg17-trixie`
+- `pg17-bookworm`, `0.8.5-pg17-bookworm`, `pg17`, `0.8.5-pg17`
+- `pg16-trixie`, `0.8.5-pg16-trixie`
+- `pg16-bookworm`, `0.8.5-pg16-bookworm`, `pg16`, `0.8.5-pg16`
+- `pg15-trixie`, `0.8.5-pg15-trixie`
+- `pg15-bookworm`, `0.8.5-pg15-bookworm`, `pg15`, `0.8.5-pg15`
+- `pg14-trixie`, `0.8.5-pg14-trixie`
+- `pg14-bookworm`, `0.8.5-pg14-bookworm`, `pg14`, `0.8.5-pg14`
+- `pg13-trixie`, `0.8.5-pg13-trixie`
+- `pg13-bookworm`, `0.8.5-pg13-bookworm`, `pg13`, `0.8.5-pg13`
 
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.8.1 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 .
 ```
@@ -1288,7 +1293,6 @@ Thanks to:
 - [k-means++: The Advantage of Careful Seeding](https://theory.stanford.edu/~sergei/papers/kMeansPP-soda.pdf)
 - [Concept Decompositions for Large Sparse Text Data using Clustering](https://www.cs.utexas.edu/users/inderjit/public_papers/concept_mlj.pdf)
 - [Efficient and Robust Approximate Nearest Neighbor Search using Hierarchical Navigable Small World Graphs](https://arxiv.org/ftp/arxiv/papers/1603/1603.09320.pdf)
-- [HQANN: Efficient and Robust Similarity Search for Hybrid Queries with Structured and Unstructured Constraints](https://arxiv.org/pdf/2207.07940.pdf)
 
 ## History
 
@@ -1335,7 +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--0.8.1--0.8.2.sql

@@ -0,0 +1,2 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION vector UPDATE TO '0.8.2'" to load this file. \quit

sql/vector--0.8.1--0.9.0.sql

@@ -1,10 +0,0 @@
--- complain if script is sourced in psql, rather than via CREATE EXTENSION
-\echo Use "ALTER EXTENSION vector UPDATE TO '0.9.0'" to load this file. \quit
-
-CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
-	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR CLASS vector_integer_ops
-	DEFAULT FOR TYPE integer USING hnsw AS
-	OPERATOR 2 = (integer, integer),
-	FUNCTION 4 hnsw_attribute_distance(integer, integer);

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

sql/vector.sql

@@ -916,13 +916,3 @@ CREATE OPERATOR CLASS sparsevec_l1_ops
 	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 l1_distance(sparsevec, sparsevec),
 	FUNCTION 3 hnsw_sparsevec_support(internal);
-
--- hnsw attributes
-
-CREATE FUNCTION hnsw_attribute_distance(integer, integer) RETURNS float8
-	AS 'MODULE_PATHNAME', 'hnsw_int4_attribute_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR CLASS vector_integer_ops
-	DEFAULT FOR TYPE integer USING hnsw AS
-	OPERATOR 2 = (integer, integer),
-	FUNCTION 4 hnsw_attribute_distance(integer, integer);

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 */
@@ -169,7 +171,7 @@ BitJaccardDistanceAvx512Popcount(uint32 bytes, unsigned char *ax, unsigned char
 #endif
 
 TARGET_XSAVE static bool
-SupportsAvx512Popcount()
+SupportsAvx512Popcount(void)
 {
 	unsigned int exx[4] = {0, 0, 0, 0};
 

src/bitvec.c

@@ -2,6 +2,7 @@
 
 #include "bitutils.h"
 #include "bitvec.h"
+#include "fmgr.h"
 #include "utils/varbit.h"
 #include "vector.h"
 

src/halfutils.c

@@ -1,5 +1,7 @@
 #include "postgres.h"
 
+#include <math.h>
+
 #include "halfutils.h"
 #include "halfvec.h"
 

src/halfvec.c

@@ -13,12 +13,20 @@
 #include "port.h"				/* for strtof() */
 #include "sparsevec.h"
 #include "utils/array.h"
-#include "utils/builtins.h"
 #include "utils/float.h"
+#include "utils/fmgrprotos.h"
 #include "utils/lsyscache.h"
-#include "utils/numeric.h"
+#include "utils/varbit.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
+#if PG_VERSION_NUM >= 170000
+#include "parser/scansup.h"
+#endif
+
 #define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
 #define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
 
@@ -129,9 +137,9 @@ InitHalfVector(int dim)
 	return result;
 }
 
-/*
- * Check for whitespace, since array_isspace() is static
- */
+#if PG_VERSION_NUM >= 170000
+#define halfvec_isspace(ch) scanner_isspace(ch)
+#else
 static inline bool
 halfvec_isspace(char ch)
 {
@@ -144,6 +152,7 @@ halfvec_isspace(char ch)
 		return true;
 	return false;
 }
+#endif
 
 /*
  * Check state array

src/hnsw.c

@@ -1,18 +1,25 @@
 #include "postgres.h"
 
 #include <float.h>
+#include <limits.h>
 #include <math.h>
 
 #include "access/amapi.h"
+#include "access/genam.h"
 #include "access/reloptions.h"
 #include "commands/progress.h"
 #include "commands/vacuum.h"
+#include "fmgr.h"
 #include "hnsw.h"
 #include "miscadmin.h"
+#include "nodes/pg_list.h"
+#include "storage/lwlock.h"
 #include "utils/float.h"
 #include "utils/guc.h"
+#include "utils/relcache.h"
 #include "utils/selfuncs.h"
 #include "utils/spccache.h"
+#include "vector.h"
 
 #if PG_VERSION_NUM < 150000
 #define MarkGUCPrefixReserved(x) EmitWarningsOnPlaceholders(x)
@@ -260,10 +267,68 @@ FUNCTION_PREFIX PG_FUNCTION_INFO_V1(hnswhandler);
 Datum
 hnswhandler(PG_FUNCTION_ARGS)
 {
+#if PG_VERSION_NUM >= 190000
+	static const IndexAmRoutine amroutine = {
+		.type = T_IndexAmRoutine,
+		.amstrategies = 0,
+		.amsupport = 3,
+		.amoptsprocnum = 0,
+		.amcanorder = false,
+		.amcanorderbyop = true,
+		.amcanhash = false,
+		.amconsistentequality = false,
+		.amconsistentordering = false,
+		.amcanbackward = false,
+		.amcanunique = false,
+		.amcanmulticol = false,
+		.amoptionalkey = true,
+		.amsearcharray = false,
+		.amsearchnulls = false,
+		.amstorage = false,
+		.amclusterable = false,
+		.ampredlocks = false,
+		.amcanparallel = false,
+		.amcanbuildparallel = true,
+		.amcaninclude = false,
+		.amusemaintenanceworkmem = false,
+		.amsummarizing = false,
+		.amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL,
+		.amkeytype = InvalidOid,
+
+		.ambuild = hnswbuild,
+		.ambuildempty = hnswbuildempty,
+		.aminsert = hnswinsert,
+		.aminsertcleanup = NULL,
+		.ambulkdelete = hnswbulkdelete,
+		.amvacuumcleanup = hnswvacuumcleanup,
+		.amcanreturn = NULL,
+		.amcostestimate = hnswcostestimate,
+		.amgettreeheight = NULL,
+		.amoptions = hnswoptions,
+		.amproperty = NULL,
+		.ambuildphasename = hnswbuildphasename,
+		.amvalidate = hnswvalidate,
+		.amadjustmembers = NULL,
+		.ambeginscan = hnswbeginscan,
+		.amrescan = hnswrescan,
+		.amgettuple = hnswgettuple,
+		.amgetbitmap = NULL,
+		.amendscan = hnswendscan,
+		.ammarkpos = NULL,
+		.amrestrpos = NULL,
+		.amestimateparallelscan = NULL,
+		.aminitparallelscan = NULL,
+		.amparallelrescan = NULL,
+		.amtranslatestrategy = NULL,
+		.amtranslatecmptype = NULL,
+	};
+
+	PG_RETURN_POINTER(&amroutine);
+#else
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 4;
+	amroutine->amsupport = 3;
 	amroutine->amoptsprocnum = 0;
 	amroutine->amcanorder = false;
 	amroutine->amcanorderbyop = true;
@@ -274,7 +339,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 #endif
 	amroutine->amcanbackward = false;	/* can change direction mid-scan */
 	amroutine->amcanunique = false;
-	amroutine->amcanmulticol = true;
+	amroutine->amcanmulticol = false;
 	amroutine->amoptionalkey = true;
 	amroutine->amsearcharray = false;
 	amroutine->amsearchnulls = false;
@@ -333,18 +398,5 @@ hnswhandler(PG_FUNCTION_ARGS)
 #endif
 
 	PG_RETURN_POINTER(amroutine);
-}
-
-/*
- * Get the distance between two int4 attributes
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(hnsw_int4_attribute_distance);
-Datum
-hnsw_int4_attribute_distance(PG_FUNCTION_ARGS)
-{
-	int32		a = PG_GETARG_INT32(0);
-	int32		b = PG_GETARG_INT32(1);
-	double		distance = ((double) a) - ((double) b);
-
-	PG_RETURN_FLOAT8(distance);
+#endif
 }

src/hnsw.h

@@ -3,15 +3,29 @@
 
 #include "postgres.h"
 
+#include <math.h>
+
 #include "access/genam.h"
 #include "access/parallel.h"
 #include "lib/pairingheap.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
+#include "storage/bufpage.h"
+#include "storage/condition_variable.h"
+#include "storage/lwlock.h"
+#include "storage/s_lock.h"
 #include "utils/relptr.h"
 #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
+
 #define HNSW_MAX_DIM 2000
 #define HNSW_MAX_NNZ 1000
 
@@ -19,7 +33,6 @@
 #define HNSW_DISTANCE_PROC 1
 #define HNSW_NORM_PROC 2
 #define HNSW_TYPE_INFO_PROC 3
-#define HNSW_ATTRIBUTE_DISTANCE_PROC 4
 
 #define HNSW_VERSION	1
 #define HNSW_MAGIC_NUMBER 0xA953A953
@@ -69,6 +82,21 @@
 #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)
@@ -108,8 +136,6 @@
 #define HnswPtrPointer(hp) (hp).ptr
 #define HnswPtrOffset(hp) relptr_offset((hp).relptr)
 
-#define HnswUseIndexTuple(index) (IndexRelationGetNumberOfAttributes(index) > 1)
-
 /* Variables */
 extern int	hnsw_ef_search;
 extern int	hnsw_iterative_scan;
@@ -137,7 +163,6 @@ HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
 HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
 HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
 HnswPtrDeclare(char, DatumRelptr, DatumPtr);
-HnswPtrDeclare(IndexTupleData, IndexTupleRelptr, IndexTuplePtr);
 
 struct HnswElementData
 {
@@ -154,7 +179,6 @@ struct HnswElementData
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
 	DatumPtr	value;
-	IndexTuplePtr itup;
 	LWLock		lock;
 };
 
@@ -180,7 +204,6 @@ typedef struct HnswSearchCandidate
 	pairingheap_node w_node;
 	HnswElementPtr element;
 	double		distance;
-	bool		matches;
 }			HnswSearchCandidate;
 
 /* HNSW index options */
@@ -259,16 +282,14 @@ typedef struct HnswTypeInfo
 
 typedef struct HnswSupport
 {
-	FmgrInfo   *procinfo[2];
+	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
-	Oid		   *collation;
+	Oid			collation;
 }			HnswSupport;
 
 typedef struct HnswQuery
 {
 	Datum		value;
-	IndexTuple	itup;
-	ScanKeyData *keyData;
 }			HnswQuery;
 
 typedef struct HnswBuildState
@@ -297,8 +318,6 @@ typedef struct HnswBuildState
 	HnswGraph  *graph;
 	double		ml;
 	int			maxLevel;
-	bool		useIndexTuple;
-	TupleDesc	tupdesc;
 
 	/* Memory */
 	MemoryContext graphCtx;
@@ -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;
@@ -427,32 +447,30 @@ 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, bool inMemory, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples);
+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 inMemory);
-HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, HnswQuery * q, Relation rel, HnswSupport * support, bool loadVec, bool inMemory);
+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);
 void		HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
 HnswNeighborArray *HnswInitNeighborArray(int lm, HnswAllocator * allocator);
 void		HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * alloc);
-bool		HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, ItemPointer heaptid, bool building, TupleDesc tupdesc);
+bool		HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building);
 void		HnswUpdateNeighborsOnDisk(Relation index, HnswSupport * support, HnswElement e, int m, bool checkExisting, bool building);
-void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index);
-void		HnswLoadElement(HnswElement element, double *distance, bool *matches, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance);
-void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple);
+void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
+void		HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance);
+bool		HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support);
+void		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
 void		HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newElement, float distance, int lm, int *updateIdx, Relation index, HnswSupport * support);
-bool		HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support, TupleDesc tupdesc);
 bool		HnswLoadNeighborTids(HnswElement element, ItemPointerData *indextids, Relation index, int m, int lm, int lc);
 void		HnswInitLockTranche(void);
 const		HnswTypeInfo *HnswGetTypeInfo(Relation index);
 PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
-Size		HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple);
-bool		HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

src/hnswbuild.c

@@ -36,11 +36,14 @@
  */
 #include "postgres.h"
 
-#include <math.h>
+#include <limits.h>
 
+#include "access/genam.h"
 #include "access/parallel.h"
+#include "access/relscan.h"
 #include "access/table.h"
 #include "access/tableam.h"
+#include "access/tupdesc.h"
 #include "access/xact.h"
 #include "access/xloginsert.h"
 #include "catalog/index.h"
@@ -48,11 +51,15 @@
 #include "commands/progress.h"
 #include "hnsw.h"
 #include "miscadmin.h"
+#include "nodes/execnodes.h"
 #include "optimizer/optimizer.h"
 #include "storage/bufmgr.h"
+#include "storage/condition_variable.h"
 #include "tcop/tcopprot.h"
 #include "utils/datum.h"
 #include "utils/memutils.h"
+#include "utils/rel.h"
+#include "utils/snapmgr.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -73,6 +80,8 @@
 #define PARALLEL_KEY_HNSW_AREA			UINT64CONST(0xA000000000000002)
 #define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000003)
 
+#define HNSW_MAX_GRAPH_MEMORY (SIZE_MAX / 2)
+
 /*
  * Create the metapage
  */
@@ -152,7 +161,6 @@ CreateGraphPages(HnswBuildState * buildstate)
 	Page		page;
 	HnswElementPtr iter = buildstate->graph->head;
 	char	   *base = buildstate->hnswarea;
-	bool		useIndexTuple = buildstate->useIndexTuple;
 
 	/* Calculate sizes */
 	maxSize = HNSW_MAX_SIZE;
@@ -172,6 +180,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
+		Pointer		valuePtr = HnswPtrAccess(base, element->value);
 
 		/* Update iterator */
 		iter = element->next;
@@ -180,7 +189,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 		MemSet(etup, 0, HNSW_TUPLE_ALLOC_SIZE);
 
 		/* Calculate sizes */
-		etupSize = HnswGetElementTupleSize(base, element, useIndexTuple);
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(valuePtr));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
@@ -190,7 +199,7 @@ CreateGraphPages(HnswBuildState * buildstate)
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 					 errmsg("index tuple too large")));
 
-		HnswSetElementTuple(base, etup, element, useIndexTuple);
+		HnswSetElementTuple(base, etup, element);
 
 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
@@ -331,18 +340,19 @@ AddDuplicateInMemory(HnswElement element, HnswElement dup)
  * Find duplicate element
  */
 static bool
-FindDuplicateInMemory(char *base, HnswElement element, bool useIndexTuple, TupleDesc tupdesc)
+FindDuplicateInMemory(char *base, HnswElement element)
 {
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);
+	Datum		value = HnswGetValue(base, element);
 
 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
+		Datum		neighborValue = HnswGetValue(base, neighborElement);
 
 		/* Exit early since ordered by distance */
-		if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
+		if (!datumIsEqual(value, neighborValue, false, -1))
 			return false;
 
 		/* Check for space */
@@ -369,7 +379,7 @@ AddElementInMemory(char *base, HnswGraph * graph, HnswElement element)
  * Update neighbors
  */
 static void
-UpdateNeighborsInMemory(char *base, Relation index, HnswSupport * support, HnswElement e, int m)
+UpdateNeighborsInMemory(char *base, HnswSupport * support, HnswElement e, int m)
 {
 	for (int lc = e->level; lc >= 0; lc--)
 	{
@@ -391,7 +401,7 @@ UpdateNeighborsInMemory(char *base, Relation index, HnswSupport * support, HnswE
 			Assert(neighborElement);
 
 			LWLockAcquire(&neighborElement->lock, LW_EXCLUSIVE);
-			HnswUpdateConnection(base, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, index, support);
+			HnswUpdateConnection(base, HnswGetNeighbors(base, neighborElement, lc), e, hc->distance, lm, NULL, NULL, support);
 			LWLockRelease(&neighborElement->lock);
 		}
 	}
@@ -407,14 +417,14 @@ UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, HnswEleme
 	char	   *base = buildstate->hnswarea;
 
 	/* Look for duplicate */
-	if (FindDuplicateInMemory(base, element, buildstate->useIndexTuple, buildstate->tupdesc))
+	if (FindDuplicateInMemory(base, element))
 		return;
 
 	/* Add element */
 	AddElementInMemory(base, graph, element);
 
 	/* Update neighbors */
-	UpdateNeighborsInMemory(base, buildstate->index, support, element, m);
+	UpdateNeighborsInMemory(base, support, element, m);
 
 	/* Update entry point if needed (already have lock) */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -427,7 +437,6 @@ UpdateGraphInMemory(HnswSupport * support, HnswElement element, int m, HnswEleme
 static void
 InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 {
-	Relation	index = buildstate->index;
 	HnswGraph  *graph = buildstate->graph;
 	HnswSupport *support = &buildstate->support;
 	HnswElement entryPoint;
@@ -461,7 +470,7 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
 	}
 
 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false, true);
+	HnswFindElementNeighbors(base, element, entryPoint, NULL, support, m, efConstruction, false);
 
 	/* Update graph in memory */
 	UpdateGraphInMemory(support, element, m, entryPoint, buildstate);
@@ -480,20 +489,22 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	HnswElement element;
 	HnswAllocator *allocator = &buildstate->allocator;
 	HnswSupport *support = &buildstate->support;
+	Size		valueSize;
+	Pointer		valuePtr;
 	LWLock	   *flushLock = &graph->flushLock;
 	char	   *base = buildstate->hnswarea;
-	TupleDesc	tupdesc = buildstate->tupdesc;
-	IndexTuple	itup;
-	Size		itupSize;
-	IndexTuple	itupShared;
-	bool		unused;
+	Datum		value;
+	Size		memoryMargin;
 
-	/* Form index tuple */
-	if (!HnswFormIndexTuple(&itup, values, isnull, buildstate->typeInfo, support, tupdesc))
+	/* Form index value */
+	if (!HnswFormIndexValue(&value, values, isnull, buildstate->typeInfo, support))
 		return false;
 
-	/* Get tuple size */
-	itupSize = IndexTupleSize(itup);
+	/* Get datum size */
+	valueSize = VARSIZE_ANY(DatumGetPointer(value));
+
+	/* In a parallel build, add a margin so allocations never fail */
+	memoryMargin = base == NULL ? 0 : 1024 * 1024;
 
 	/* Ensure graph not flushed when inserting */
 	LWLockAcquire(flushLock, LW_SHARED);
@@ -503,7 +514,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	{
 		LWLockRelease(flushLock);
 
-		return HnswInsertTupleOnDisk(index, support, itup, heaptid, true, tupdesc);
+		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
 	}
 
 	/*
@@ -516,7 +527,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	 * Check that we have enough memory available for the new element now that
 	 * we have the allocator lock, and flush pages if needed.
 	 */
-	if (graph->memoryUsed >= graph->memoryTotal)
+	if (graph->memoryUsed + memoryMargin >= graph->memoryTotal)
 	{
 		LWLockRelease(&graph->allocatorLock);
 
@@ -535,12 +546,12 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 
 		LWLockRelease(flushLock);
 
-		return HnswInsertTupleOnDisk(index, support, itup, heaptid, true, tupdesc);
+		return HnswInsertTupleOnDisk(index, support, value, heaptid, true);
 	}
 
 	/* Ok, we can proceed to allocate the element */
 	element = HnswInitElement(base, heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel, allocator);
-	itupShared = HnswAlloc(allocator, itupSize);
+	valuePtr = HnswAlloc(allocator, valueSize);
 
 	/*
 	 * We have now allocated the space needed for the element, so we don't
@@ -549,10 +560,9 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	 */
 	LWLockRelease(&graph->allocatorLock);
 
-	/* Copy the tuple */
-	memcpy(itupShared, itup, itupSize);
-	HnswPtrStore(base, element->itup, itupShared);
-	HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itupShared, 1, tupdesc, &unused)));
+	/* Copy the datum */
+	memcpy(valuePtr, DatumGetPointer(value), valueSize);
+	HnswPtrStore(base, element->value, (char *) valuePtr);
 
 	/* Create a lock for the element */
 	LWLockInitialize(&element->lock, hnsw_lock_tranche_id);
@@ -610,7 +620,7 @@ InitGraph(HnswGraph * graph, char *base, Size memoryTotal)
 	HnswPtrStore(base, graph->head, (HnswElement) NULL);
 	HnswPtrStore(base, graph->entryPoint, (HnswElement) NULL);
 	graph->memoryUsed = 0;
-	graph->memoryTotal = memoryTotal;
+	graph->memoryTotal = Min(memoryTotal, HNSW_MAX_GRAPH_MEMORY);
 	graph->flushed = false;
 	graph->indtuples = 0;
 	SpinLockInit(&graph->lock);
@@ -651,9 +661,17 @@ static void *
 HnswSharedMemoryAlloc(Size size, void *state)
 {
 	HnswBuildState *buildstate = (HnswBuildState *) state;
-	void	   *chunk = buildstate->hnswarea + buildstate->graph->memoryUsed;
+	Size		alignedSize = MAXALIGN(size);
+	void	   *chunk;
 
-	buildstate->graph->memoryUsed += MAXALIGN(size);
+	if (alignedSize > 1024 * 1024)
+		elog(ERROR, "hnsw allocation too large");
+
+	if (buildstate->graph->memoryUsed + alignedSize > buildstate->graph->memoryTotal)
+		elog(ERROR, "hnsw allocator out of memory");
+
+	chunk = buildstate->hnswarea + buildstate->graph->memoryUsed;
+	buildstate->graph->memoryUsed += alignedSize;
 	return chunk;
 }
 
@@ -679,19 +697,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 				 errmsg("type not supported for hnsw index")));
 
-	/* TODO See if needed */
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 2)
-		elog(ERROR, "index cannot have more than two columns");
-
-	if (!OidIsValid(index_getprocid(index, 1, HNSW_DISTANCE_PROC)))
-		elog(ERROR, "first column must be a vector");
-
-	for (int i = 1; i < IndexRelationGetNumberOfKeyAttributes(index); i++)
-	{
-		if (!OidIsValid(index_getprocid(index, i + 1, HNSW_ATTRIBUTE_DISTANCE_PROC)))
-			elog(ERROR, "column %d cannot be a vector", i + 1);
-	}
-
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)
 		ereport(ERROR,
@@ -714,12 +719,10 @@ 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);
-	buildstate->useIndexTuple = HnswUseIndexTuple(index);
-	buildstate->tupdesc = RelationGetDescr(index);
 
 	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
 												   "Hnsw build graph context",
@@ -800,7 +803,11 @@ HnswParallelScanAndInsert(Relation heapRel, Relation indexRel, HnswShared * hnsw
 	buildstate.hnswarea = hnswarea;
 	InitAllocator(&buildstate.allocator, &HnswSharedMemoryAlloc, &buildstate);
 	scan = table_beginscan_parallel(heapRel,
-									ParallelTableScanFromHnswShared(hnswshared));
+									ParallelTableScanFromHnswShared(hnswshared)
+#if PG_VERSION_NUM >= 190000
+									,SO_NONE
+#endif
+		);
 	reltuples = table_index_build_scan(heapRel, indexRel, indexInfo,
 									   true, progress, BuildCallback,
 									   (void *) &buildstate, scan);
@@ -949,11 +956,13 @@ 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;
 
+	esthnswarea = Min(esthnswarea, HNSW_MAX_GRAPH_MEMORY);
+
 	shm_toc_estimate_chunk(&pcxt->estimator, esthnswarea);
 	shm_toc_estimate_keys(&pcxt->estimator, 2);
 
@@ -996,8 +1005,7 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 								  snapshot);
 
 	hnswarea = (char *) shm_toc_allocate(pcxt->toc, esthnswarea);
-	/* Report less than allocated so never fails */
-	InitGraph(&hnswshared->graphData, hnswarea, esthnswarea - 1024 * 1024);
+	InitGraph(&hnswshared->graphData, hnswarea, esthnswarea);
 
 	/*
 	 * Avoid base address for relptr for Postgres < 14.5

src/hnswinsert.c

@@ -1,13 +1,15 @@
 #include "postgres.h"
 
-#include <math.h>
-
+#include "access/genam.h"
 #include "access/generic_xlog.h"
 #include "hnsw.h"
+#include "nodes/execnodes.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "storage/lwlock.h"
 #include "utils/datum.h"
 #include "utils/memutils.h"
+#include "utils/rel.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -160,10 +162,9 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 	BlockNumber newInsertPage = InvalidBlockNumber;
 	uint8		tupleVersion;
 	char	   *base = NULL;
-	bool		useIndexTuple = HnswUseIndexTuple(index);
 
 	/* Calculate sizes */
-	etupSize = HnswGetElementTupleSize(base, e, useIndexTuple);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(HnswPtrAccess(base, e->value)));
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -171,7 +172,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
 
 	/* Prepare element tuple */
 	etup = palloc0(etupSize);
-	HnswSetElementTuple(base, etup, e, useIndexTuple);
+	HnswSetElementTuple(base, etup, e);
 
 	/* Prepare neighbor tuple */
 	ntup = palloc0(ntupSize);
@@ -388,9 +389,8 @@ LoadElementsForInsert(HnswNeighborArray * neighbors, HnswQuery * q, int *idx, Re
 		HnswCandidate *hc = &neighbors->items[i];
 		HnswElement element = HnswPtrAccess(base, hc->element);
 		double		distance;
-		bool		matches;
 
-		HnswLoadElement(element, &distance, &matches, q, index, support, true, NULL);
+		HnswLoadElement(element, &distance, q, index, support, true, NULL);
 		hc->distance = distance;
 
 		/* Prune element if being deleted */
@@ -434,8 +434,6 @@ GetUpdateIndex(HnswElement element, HnswElement newElement, float distance, int
 		HnswQuery	q;
 
 		q.value = HnswGetValue(base, element);
-		q.itup = HnswPtrAccess(base, element->itup);
-		q.keyData = NULL;
 
 		LoadElementsForInsert(neighbors, &q, &idx, index, support);
 
@@ -641,30 +639,21 @@ AddDuplicateOnDisk(Relation index, HnswElement element, HnswElement dup, bool bu
  * Find duplicate element
  */
 static bool
-FindDuplicateOnDisk(Relation index, HnswElement element, bool building, TupleDesc tupdesc)
+FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
 {
 	char	   *base = NULL;
 	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
 	Datum		value = HnswGetValue(base, element);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);
 
 	for (int i = 0; i < neighbors->length; i++)
 	{
 		HnswCandidate *neighbor = &neighbors->items[i];
 		HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
+		Datum		neighborValue = HnswGetValue(base, neighborElement);
 
-		if (HnswUseIndexTuple(index))
-		{
-			/* Exit early since ordered by distance */
-			if (!HnswIndexTupleIsEqual(itup, HnswPtrAccess(base, neighborElement->itup), tupdesc))
-				return false;
-		}
-		else
-		{
-			/* Exit early since ordered by distance */
-			if (!datumIsEqual(value, HnswGetValue(base, neighborElement), false, -1))
-				return false;
-		}
+		/* Exit early since ordered by distance */
+		if (!datumIsEqual(value, neighborValue, false, -1))
+			return false;
 
 		if (AddDuplicateOnDisk(index, element, neighborElement, building))
 			return true;
@@ -677,12 +666,12 @@ FindDuplicateOnDisk(Relation index, HnswElement element, bool building, TupleDes
  * Update graph on disk
  */
 static void
-UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, int m, HnswElement entryPoint, bool building, TupleDesc tupdesc)
+UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, int m, HnswElement entryPoint, bool building)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Look for duplicate */
-	if (FindDuplicateOnDisk(index, element, building, tupdesc))
+	if (FindDuplicateOnDisk(index, element, building))
 		return;
 
 	/* Add element */
@@ -704,7 +693,7 @@ UpdateGraphOnDisk(Relation index, HnswSupport * support, HnswElement element, in
  * Insert a tuple into the index
  */
 bool
-HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, ItemPointer heaptid, bool building, TupleDesc tupdesc)
+HnswInsertTupleOnDisk(Relation index, HnswSupport * support, Datum value, ItemPointer heaptid, bool building)
 {
 	HnswElement entryPoint;
 	HnswElement element;
@@ -712,7 +701,6 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, It
 	int			efConstruction = HnswGetEfConstruction(index);
 	LOCKMODE	lockmode = ShareLock;
 	char	   *base = NULL;
-	bool		unused;
 
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -726,8 +714,7 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, It
 
 	/* Create an element */
 	element = HnswInitElement(base, heaptid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
-	HnswPtrStore(base, element->itup, itup);
-	HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
+	HnswPtrStore(base, element->value, (char *) DatumGetPointer(value));
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -744,10 +731,10 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, It
 	}
 
 	/* Find neighbors for element */
-	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, false);
 
 	/* Update graph on disk */
-	UpdateGraphOnDisk(index, support, element, m, entryPoint, building, tupdesc);
+	UpdateGraphOnDisk(index, support, element, m, entryPoint, building);
 
 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -761,18 +748,17 @@ HnswInsertTupleOnDisk(Relation index, HnswSupport * support, IndexTuple itup, It
 static void
 HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid)
 {
-	IndexTuple	itup;
+	Datum		value;
 	const		HnswTypeInfo *typeInfo = HnswGetTypeInfo(index);
-	TupleDesc	tupdesc = RelationGetDescr(index);
 	HnswSupport support;
 
 	HnswInitSupport(&support, index);
 
-	/* Form index tuple */
-	if (!HnswFormIndexTuple(&itup, values, isnull, typeInfo, &support, tupdesc))
+	/* Form index value */
+	if (!HnswFormIndexValue(&value, values, isnull, typeInfo, &support))
 		return;
 
-	HnswInsertTupleOnDisk(index, &support, itup, heaptid, false, tupdesc);
+	HnswInsertTupleOnDisk(index, &support, value, heaptid, false);
 }
 
 /*

src/hnswscan.c

@@ -1,12 +1,23 @@
 #include "postgres.h"
 
+#include <limits.h>
+
+#include "access/genam.h"
 #include "access/relscan.h"
 #include "hnsw.h"
+#include "lib/pairingheap.h"
+#include "miscadmin.h"
+#include "nodes/pg_list.h"
 #include "pgstat.h"
-#include "storage/bufmgr.h"
 #include "storage/lmgr.h"
 #include "utils/float.h"
 #include "utils/memutils.h"
+#include "utils/relcache.h"
+#include "utils/snapmgr.h"
+
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
 
 /*
  * Algorithm 5 from paper
@@ -22,30 +33,26 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	int			m;
 	HnswElement entryPoint;
 	char	   *base = NULL;
-	bool		inMemory = false;
 	HnswQuery  *q = &so->q;
 
-	q->value = value;
-	q->itup = NULL;
-	q->keyData = scan->keyData;
-
 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);
 
+	q->value = value;
 	so->m = m;
 
 	if (entryPoint == NULL)
 		return NIL;
 
-	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, support, false, inMemory));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, support, false));
 
 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(base, q, ep, 1, lc, index, support, m, false, NULL, inMemory, NULL, NULL, true, NULL);
+		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, inMemory, &so->v, hnsw_iterative_scan != HNSW_ITERATIVE_SCAN_OFF ? &so->discarded : NULL, true, &so->tuples);
+	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);
 }
 
 /*
@@ -76,7 +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, false, &so->v, &so->discarded, false, &so->tuples);
+	return HnswSearchLayer(base, &so->q, ep, batch_size, 0, index, &so->support, so->m, false, NULL, &so->v, &so->discarded, false, &so->tuples);
 }
 
 /*
@@ -100,7 +107,7 @@ GetScanValue(IndexScanDesc scan)
 
 		/* Normalize if needed */
 		if (so->support.normprocinfo != NULL)
-			value = HnswNormValue(so->typeInfo, so->support.collation[0], value);
+			value = HnswNormValue(so->typeInfo, so->support.collation, value);
 	}
 
 	return value;
@@ -146,7 +153,7 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	/* Calculate max memory */
 	/* Add 256 extra bytes to fill last block when close */
 	maxMemory = (double) work_mem * hnsw_scan_mem_multiplier * 1024.0 + 256;
-	so->maxMemory = Min(maxMemory, (double) SIZE_MAX);
+	so->maxMemory = Min(maxMemory, (double) (SIZE_MAX / 2));
 
 	scan->opaque = so;
 
@@ -287,7 +294,7 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		element = HnswPtrAccess(base, sc->element);
 
 		/* Move to next element if no valid heap TIDs */
-		if (!sc->matches || element->heaptidsLength == 0)
+		if (element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
 

src/hnswutils.c

@@ -2,18 +2,20 @@
 
 #include <math.h>
 
+#include "access/genam.h"
 #include "access/generic_xlog.h"
-#include "catalog/pg_type.h"
-#include "catalog/pg_type_d.h"
 #include "common/hashfn.h"
 #include "fmgr.h"
 #include "hnsw.h"
 #include "lib/pairingheap.h"
+#include "nodes/pg_list.h"
+#include "port/atomics.h"
 #include "sparsevec.h"
 #include "storage/bufmgr.h"
 #include "utils/datum.h"
 #include "utils/memdebug.h"
 #include "utils/rel.h"
+#include "vector.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -150,39 +152,11 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 void
 HnswInitSupport(HnswSupport * support, Relation index)
 {
-	support->procinfo[0] = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
-
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
-		support->procinfo[1] = index_getprocinfo(index, 2, HNSW_ATTRIBUTE_DISTANCE_PROC);
-
-	support->collation = index->rd_indcollation;
+	support->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
+	support->collation = index->rd_indcollation[0];
 	support->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 }
 
-/*
- * Get element tuple size
- */
-Size
-HnswGetElementTupleSize(char *base, HnswElement element, bool useIndexTuple)
-{
-	Size		size;
-
-	if (useIndexTuple)
-	{
-		IndexTuple	itup = HnswPtrAccess(base, element->itup);
-
-		size = IndexTupleSize(itup);
-	}
-	else
-	{
-		Pointer		valuePtr = HnswPtrAccess(base, element->value);
-
-		size = VARSIZE_ANY(valuePtr);
-	}
-
-	return HNSW_ELEMENT_TUPLE_SIZE(size);
-}
-
 /*
  * Normalize value
  */
@@ -198,38 +172,7 @@ HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value)
 bool
 HnswCheckNorm(HnswSupport * support, Datum value)
 {
-	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation[0], value)) > 0;
-}
-
-/*
- * Check if index tuples are equal
- */
-bool
-HnswIndexTupleIsEqual(IndexTuple a, IndexTuple b, TupleDesc tupdesc)
-{
-	for (int i = 0; i < tupdesc->natts; i++)
-	{
-		bool		nullA;
-		bool		nullB;
-
-		Datum		datumA = index_getattr(a, i + 1, tupdesc, &nullA);
-		Datum		datumB = index_getattr(b, i + 1, tupdesc, &nullB);
-
-		if (nullA || nullB)
-		{
-			if (nullA != nullB)
-				return false;
-		}
-		else
-		{
-			Form_pg_attribute att = TupleDescAttr(tupdesc, i);
-
-			if (!datumIsEqual(datumA, datumB, att->attbyval, att->attlen))
-				return false;
-		}
-	}
-
-	return true;
+	return DatumGetFloat8(FunctionCall1Coll(support->normprocinfo, support->collation, value)) > 0;
 }
 
 /*
@@ -319,8 +262,7 @@ HnswInitElement(char *base, ItemPointer heaptid, int m, double ml, int maxLevel,
 
 	HnswInitNeighbors(base, element, m, allocator);
 
-	HnswPtrStore(base, element->value, (Pointer) NULL);
-	HnswPtrStore(base, element->itup, (IndexTuple) NULL);
+	HnswPtrStore(base, element->value, (char *) NULL);
 
 	return element;
 }
@@ -346,8 +288,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	HnswPtrStore(base, element->neighbors, (HnswNeighborArrayPtr *) NULL);
-	HnswPtrStore(base, element->value, (Pointer) NULL);
-	HnswPtrStore(base, element->itup, (IndexTuple) NULL);
+	HnswPtrStore(base, element->value, (char *) NULL);
 	return element;
 }
 
@@ -460,13 +401,11 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 }
 
 /*
- * Form index tuple
+ * Form index value
  */
 bool
-HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support, TupleDesc tupdesc)
+HnswFormIndexValue(Datum *out, Datum *values, bool *isnull, const HnswTypeInfo * typeInfo, HnswSupport * support)
 {
-	Datum		newValues[2];
-
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
@@ -480,14 +419,10 @@ HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeI
 		if (!HnswCheckNorm(support, value))
 			return false;
 
-		value = HnswNormValue(typeInfo, support->collation[0], value);
+		value = HnswNormValue(typeInfo, support->collation, value);
 	}
 
-	newValues[0] = value;
-	for (int i = 1; i < tupdesc->natts; i++)
-		newValues[i] = values[i];
-
-	*out = index_form_tuple(tupdesc, newValues, isnull);
+	*out = value;
 
 	return true;
 }
@@ -496,8 +431,10 @@ HnswFormIndexTuple(IndexTuple *out, Datum *values, bool *isnull, const HnswTypeI
  * Set element tuple, except for neighbor info
  */
 void
-HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple)
+HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
 {
+	Pointer		valuePtr = HnswPtrAccess(base, element->value);
+
 	etup->type = HNSW_ELEMENT_TUPLE_TYPE;
 	etup->level = element->level;
 	etup->deleted = 0;
@@ -509,19 +446,7 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-
-	if (useIndexTuple)
-	{
-		IndexTuple	itup = HnswPtrAccess(base, element->itup);
-
-		memcpy(&etup->data, itup, IndexTupleSize(itup));
-	}
-	else
-	{
-		Pointer		valuePtr = HnswPtrAccess(base, element->value);
-
-		memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
-	}
+	memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
 }
 
 /*
@@ -563,7 +488,7 @@ HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
  * Load an element from a tuple
  */
 void
-HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index)
+HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec)
 {
 	element->level = etup->level;
 	element->deleted = etup->deleted;
@@ -587,128 +512,26 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	if (loadVec)
 	{
 		char	   *base = NULL;
+		Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
 
-		if (HnswUseIndexTuple(index))
-		{
-			IndexTuple	itup = CopyIndexTuple((IndexTuple) &etup->data);
-			TupleDesc	tupdesc = RelationGetDescr(index);
-			bool		unused;
-
-			HnswPtrStore(base, element->itup, itup);
-			HnswPtrStore(base, element->value, DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused)));
-		}
-		else
-		{
-			Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
-
-			HnswPtrStore(base, element->value, DatumGetPointer(value));
-		}
+		HnswPtrStore(base, element->value, (char *) DatumGetPointer(value));
 	}
 }
 
-/*
- * Get the attribute distance
- */
-static inline double
-AttributeDistance(double e)
-{
-	/* TODO Better bias */
-	/* must be >> max(w * g) + 1 / log10(2) */
-	double		bias = 4.32;
-
-	return e > 0 ? bias - 1.0 / log10(e + 1) : 0;
-}
-
 /*
  * Calculate the distance between values
  */
-static double
-HnswGetDistance(IndexTuple itup, Datum vec, HnswQuery * q, Relation index, HnswSupport * support, bool *matches)
+static inline double
+HnswGetDistance(Datum a, Datum b, HnswSupport * support)
 {
-	double		g;
-
-	if (DatumGetPointer(q->value) == NULL)
-		g = 0;
-	else
-		g = DatumGetFloat8(FunctionCall2Coll(support->procinfo[0], support->collation[0], q->value, vec));
-
-	Assert(PointerIsValid(matches));
-	*matches = true;
-
-	if (IndexRelationGetNumberOfKeyAttributes(index) > 1)
-	{
-		double		w = 0.25;
-		double		e = 0.0;
-		TupleDesc	tupdesc = RelationGetDescr(index);
-
-		if (q->keyData)
-		{
-			/* TODO need to pass length of key data */
-			int			keyCount = 1;
-
-			for (int i = 0; i < keyCount; i++)
-			{
-				ScanKey		key = &q->keyData[i];
-				bool		isnull;
-				Datum		value = index_getattr(itup, key->sk_attno, tupdesc, &isnull);
-				bool		attnull = key->sk_flags & SK_ISNULL;
-
-				if (isnull || attnull)
-				{
-					if (isnull != attnull)
-					{
-						e += 1000;
-						*matches = false;
-					}
-				}
-				else if (!DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation, value, key->sk_argument)))
-				{
-					double		ei = fabs(DatumGetFloat8(FunctionCall2Coll(support->procinfo[key->sk_attno - 1], support->collation[key->sk_attno - 1], value, key->sk_argument)));
-
-					if (ei > 0)
-						e += ei;
-					else
-						/* Distance is zero for inequality */
-						e += 1000;
-
-					*matches = false;
-				}
-			}
-
-			return w * g + AttributeDistance(e);
-		}
-		else if (q->itup)
-		{
-			int			keyCount = IndexRelationGetNumberOfKeyAttributes(index) - 1;
-
-			for (int i = 0; i < keyCount; i++)
-			{
-				bool		isnull;
-				bool		attnull;
-				Datum		value = index_getattr(itup, i + 2, tupdesc, &isnull);
-				Datum		value2 = index_getattr(q->itup, i + 2, tupdesc, &attnull);
-
-				if (isnull || attnull)
-				{
-					if (isnull != attnull)
-						e += 1000;
-				}
-				else
-					e += fabs(DatumGetFloat8(FunctionCall2Coll(support->procinfo[i + 1], support->collation[i + 1], value, value2)));
-			}
-
-			return w * g + AttributeDistance(e);
-		}
-	}
-
-	return g;
+	return DatumGetFloat8(FunctionCall2Coll(support->procinfo, support->collation, a, b));
 }
 
 /*
  * Load an element and optionally get its distance from q
  */
 static void
-HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, bool *matches, 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;
@@ -723,26 +546,16 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
 
 	Assert(HnswIsElementTuple(etup));
 
+	if (unlikely(etup->deleted))
+		elog(ERROR, "cannot load deleted element");
+
 	/* Calculate distance */
 	if (distance != NULL)
 	{
-		IndexTuple	itup = NULL;
-		Datum		value;
-
-		if (HnswUseIndexTuple(index))
-		{
-			TupleDesc	tupdesc = RelationGetDescr(index);
-			bool		unused;
-
-			itup = (IndexTuple) &etup->data;
-			value = index_getattr(itup, 1, tupdesc, &unused);
-		}
+		if (DatumGetPointer(q->value) == NULL)
+			*distance = 0;
 		else
-		{
-			value = PointerGetDatum(&etup->data);
-		}
-
-		*distance = HnswGetDistance(itup, value, q, index, support, matches);
+			*distance = HnswGetDistance(q->value, PointerGetDatum(&etup->data), support);
 	}
 
 	/* Load element */
@@ -751,7 +564,7 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
 		if (*element == NULL)
 			*element = HnswInitElementFromBlock(blkno, offno);
 
-		HnswLoadElementFromTuple(*element, etup, true, loadVec, index);
+		HnswLoadElementFromTuple(*element, etup, true, loadVec);
 	}
 
 	UnlockReleaseBuffer(buf);
@@ -761,34 +574,32 @@ HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, double *distance, boo
  * Load an element and optionally get its distance from q
  */
 void
-HnswLoadElement(HnswElement element, double *distance, bool *matches, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
+HnswLoadElement(HnswElement element, double *distance, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, double *maxDistance)
 {
-	HnswLoadElementImpl(element->blkno, element->offno, distance, matches, q, index, support, loadVec, maxDistance, &element);
+	HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, support, loadVec, maxDistance, &element);
 }
 
 /*
  * Get the distance for an element
  */
 static double
-GetElementDistance(char *base, HnswElement element, bool *matches, HnswQuery * q, Relation index, HnswSupport * support)
+GetElementDistance(char *base, HnswElement element, HnswQuery * q, HnswSupport * support)
 {
 	Datum		value = HnswGetValue(base, element);
-	IndexTuple	itup = HnswPtrAccess(base, element->itup);
 
-	return HnswGetDistance(itup, value, q, index, support, matches);
+	return HnswGetDistance(q->value, value, support);
 }
 
 /*
  * Allocate a search candidate
  */
 static HnswSearchCandidate *
-HnswInitSearchCandidate(char *base, HnswElement element, double distance, bool matches)
+HnswInitSearchCandidate(char *base, HnswElement element, double distance)
 {
 	HnswSearchCandidate *sc = palloc(sizeof(HnswSearchCandidate));
 
 	HnswPtrStore(base, sc->element, element);
 	sc->distance = distance;
-	sc->matches = matches;
 	return sc;
 }
 
@@ -796,17 +607,17 @@ HnswInitSearchCandidate(char *base, HnswElement element, double distance, bool m
  * Create a candidate for the entry point
  */
 HnswSearchCandidate *
-HnswEntryCandidate(char *base, HnswElement entryPoint, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec, bool inMemory)
+HnswEntryCandidate(char *base, HnswElement entryPoint, HnswQuery * q, Relation index, HnswSupport * support, bool loadVec)
 {
+	bool		inMemory = index == NULL;
 	double		distance;
-	bool		matches;
 
 	if (inMemory)
-		distance = GetElementDistance(base, entryPoint, &matches, q, index, support);
+		distance = GetElementDistance(base, entryPoint, q, support);
 	else
-		HnswLoadElement(entryPoint, &distance, &matches, q, index, support, loadVec, NULL);
+		HnswLoadElement(entryPoint, &distance, q, index, support, loadVec, NULL);
 
-	return HnswInitSearchCandidate(base, entryPoint, distance, matches);
+	return HnswInitSearchCandidate(base, entryPoint, distance);
 }
 
 /*
@@ -1009,7 +820,7 @@ HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *u
  * 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, bool inMemory, visited_hash * v, pairingheap **discarded, bool initVisited, int64 *tuples)
+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);
@@ -1022,8 +833,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 	int			lm = HnswGetLayerM(m, lc);
 	HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
 	int			unvisitedLength;
-	uint64		additional = 0;
-	uint64		maxAdditional = q->keyData && lc == 0 ? 10000 : 0;
+	bool		inMemory = index == NULL;
 
 	if (v == NULL)
 	{
@@ -1064,10 +874,6 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 		pairingheap_add(C, &sc->c_node);
 		pairingheap_add(W, &sc->w_node);
 
-		/* Do not count elements that do not match filter towards ef */
-		if (!sc->matches && ++additional <= maxAdditional)
-			continue;
-
 		/*
 		 * Do not count elements being deleted towards ef when vacuuming. It
 		 * would be ideal to do this for inserts as well, but this could
@@ -1102,7 +908,6 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			HnswElement eElement;
 			HnswSearchCandidate *e;
 			double		eDistance;
-			bool		eMatches;
 			bool		alwaysAdd = wlen < ef;
 
 			f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
@@ -1110,7 +915,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			if (inMemory)
 			{
 				eElement = unvisited[i].element;
-				eDistance = GetElementDistance(base, eElement, &eMatches, q, index, support);
+				eDistance = GetElementDistance(base, eElement, q, support);
 			}
 			else
 			{
@@ -1120,18 +925,18 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 
 				/* Avoid any allocations if not adding */
 				eElement = NULL;
-				HnswLoadElementImpl(blkno, offno, &eDistance, &eMatches, 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;
 			}
 
-			if (eElement == NULL || !(eDistance < f->distance || alwaysAdd))
+			if (!(eDistance < f->distance || alwaysAdd))
 			{
 				if (discarded != NULL)
 				{
 					/* Create a new candidate */
-					e = HnswInitSearchCandidate(base, eElement, eDistance, eMatches);
+					e = HnswInitSearchCandidate(base, eElement, eDistance);
 					pairingheap_add(*discarded, &e->w_node);
 				}
 
@@ -1143,7 +948,7 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 				continue;
 
 			/* Create a new candidate */
-			e = HnswInitSearchCandidate(base, eElement, eDistance, eMatches);
+			e = HnswInitSearchCandidate(base, eElement, eDistance);
 			pairingheap_add(C, &e->c_node);
 			pairingheap_add(W, &e->w_node);
 
@@ -1154,10 +959,6 @@ HnswSearchLayer(char *base, HnswQuery * q, List *ep, int ef, int lc, Relation in
 			 */
 			if (CountElement(skipElement, eElement))
 			{
-				/* Do not count elements that do not match filter towards ef */
-				if (!e->matches && ++additional <= maxAdditional)
-					continue;
-
 				wlen++;
 
 				/* No need to decrement wlen */
@@ -1235,24 +1036,18 @@ CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
  * Check if an element is closer to q than any element from R
  */
 static bool
-CheckElementCloser(char *base, HnswCandidate * e, List *r, Relation index, HnswSupport * support)
+CheckElementCloser(char *base, HnswCandidate * e, List *r, HnswSupport * support)
 {
 	HnswElement eElement = HnswPtrAccess(base, e->element);
-	HnswQuery	q;
+	Datum		eValue = HnswGetValue(base, eElement);
 	ListCell   *lc2;
 
-	q.value = HnswGetValue(base, eElement);
-	q.itup = HnswPtrAccess(base, eElement->itup);
-	q.keyData = NULL;
-
 	foreach(lc2, r)
 	{
 		HnswCandidate *ri = lfirst(lc2);
 		HnswElement riElement = HnswPtrAccess(base, ri->element);
 		Datum		riValue = HnswGetValue(base, riElement);
-		IndexTuple	ritup = HnswPtrAccess(base, riElement->itup);
-		bool		matches;
-		float		distance = HnswGetDistance(ritup, riValue, &q, index, support, &matches);
+		float		distance = HnswGetDistance(eValue, riValue, support);
 
 		if (distance <= e->distance)
 			return false;
@@ -1265,7 +1060,7 @@ CheckElementCloser(char *base, HnswCandidate * e, List *r, Relation index, HnswS
  * Algorithm 4 from paper
  */
 static List *
-SelectNeighbors(char *base, List *c, int lm, Relation index, HnswSupport * support, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
+SelectNeighbors(char *base, List *c, int lm, HnswSupport * support, bool *closerSet, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
@@ -1299,7 +1094,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, HnswSupport * suppo
 
 		/* Use previous state of r and wd to skip work when possible */
 		if (mustCalculate)
-			e->closer = CheckElementCloser(base, e, r, index, support);
+			e->closer = CheckElementCloser(base, e, r, support);
 		else if (list_length(added) > 0)
 		{
 			/* Keep Valgrind happy for in-memory, parallel builds */
@@ -1312,7 +1107,8 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, HnswSupport * suppo
 			 */
 			if (e->closer)
 			{
-				e->closer = CheckElementCloser(base, e, added, index, support);
+				e->closer = CheckElementCloser(base, e, added, support);
+
 				if (!e->closer)
 					removedAny = true;
 			}
@@ -1324,7 +1120,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, HnswSupport * suppo
 				 */
 				if (removedAny)
 				{
-					e->closer = CheckElementCloser(base, e, r, index, support);
+					e->closer = CheckElementCloser(base, e, r, support);
 					if (e->closer)
 						added = lappend(added, e);
 				}
@@ -1332,7 +1128,7 @@ SelectNeighbors(char *base, List *c, int lm, Relation index, HnswSupport * suppo
 		}
 		else if (e == newCandidate)
 		{
-			e->closer = CheckElementCloser(base, e, r, index, support);
+			e->closer = CheckElementCloser(base, e, r, support);
 			if (e->closer)
 				added = lappend(added, e);
 		}
@@ -1409,7 +1205,7 @@ HnswUpdateConnection(char *base, HnswNeighborArray * neighbors, HnswElement newE
 			c = lappend(c, &neighbors->items[i]);
 		c = lappend(c, &newHc);
 
-		SelectNeighbors(base, c, lm, index, support, &neighbors->closerSet, &newHc, &pruned, true);
+		SelectNeighbors(base, c, lm, support, &neighbors->closerSet, &newHc, &pruned, true);
 
 		/* Should not happen */
 		if (pruned == NULL)
@@ -1480,19 +1276,17 @@ 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 inMemory)
+HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, HnswSupport * support, int m, int efConstruction, bool existing)
 {
 	List	   *ep;
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
 	HnswQuery	q;
-
 	HnswElement skipElement = existing ? element : NULL;
+	bool		inMemory = index == NULL;
 
 	q.value = HnswGetValue(base, element);
-	q.itup = HnswPtrAccess(base, element->itup);
-	q.keyData = NULL;
 
 	/* Precompute hash */
 	if (inMemory)
@@ -1503,13 +1297,13 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		return;
 
 	/* Get entry point and level */
-	ep = list_make1(HnswEntryCandidate(base, entryPoint, &q, index, support, true, inMemory));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, &q, index, support, true));
 	entryLevel = entryPoint->level;
 
 	/* 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, inMemory, NULL, NULL, true, NULL);
+		w = HnswSearchLayer(base, &q, ep, 1, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
 		ep = w;
 	}
 
@@ -1528,7 +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, inMemory, NULL, NULL, true, NULL);
+		w = HnswSearchLayer(base, &q, ep, efConstruction, lc, index, support, m, true, skipElement, NULL, NULL, true, NULL);
 
 		/* Convert search candidates to candidates */
 		foreach(lc2, w)
@@ -1552,7 +1346,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		 * sortCandidates to true for in-memory builds to enable closer
 		 * caching, but there does not seem to be a difference in performance.
 		 */
-		neighbors = SelectNeighbors(base, lw, lm, index, support, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);
+		neighbors = SelectNeighbors(base, lw, lm, support, &HnswGetNeighbors(base, element, lc)->closerSet, NULL, NULL, false);
 
 		AddConnections(base, element, neighbors, lc);
 

src/hnswvacuum.c

@@ -1,13 +1,14 @@
 #include "postgres.h"
 
-#include <math.h>
-
+#include "access/genam.h"
 #include "access/generic_xlog.h"
 #include "commands/vacuum.h"
 #include "hnsw.h"
+#include "nodes/pg_list.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
 #include "utils/memutils.h"
+#include "utils/rel.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -18,12 +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;
 }
 
 /*
@@ -36,17 +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))
 	{
@@ -76,6 +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++)
@@ -109,23 +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;
@@ -137,6 +166,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 		UnlockReleaseBuffer(buf);
 	}
+
+#ifdef HNSW_MEMORY
+	elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(CurrentMemoryContext, true) / 1024);
+#endif
 }
 
 /*
@@ -167,8 +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;
@@ -177,7 +210,8 @@ 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)
 		needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
 
 	UnlockReleaseBuffer(buf);
@@ -212,7 +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, false);
+	HnswFindElementNeighbors(base, element, entryPoint, index, support, m, efConstruction, true);
 
 	/* Zero memory for each element */
 	MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
@@ -263,12 +297,27 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 		/* Get a shared lock */
 		LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
 
-		/* Load element */
-		HnswLoadElement(highestPoint, NULL, 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);
@@ -286,7 +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
@@ -302,7 +351,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 			 * is outdated, this can remove connections at higher levels in
 			 * the graph until they are repaired, but this should be fine.
 			 */
-			HnswLoadElement(entryPoint, NULL, NULL, NULL, index, support, true, NULL);
+			HnswLoadElement(entryPoint, NULL, NULL, index, support, true, NULL);
 
 			if (NeedsUpdated(vacuumstate, entryPoint))
 			{
@@ -372,13 +421,17 @@ 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;
 
 			/* Create an element */
 			element = HnswInitElementFromBlock(blkno, offno);
-			HnswLoadElementFromTuple(element, etup, false, true, index);
+			HnswLoadElementFromTuple(element, etup, false, true);
 
 			elements = lappend(elements, element);
 		}
@@ -435,6 +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);
 	}
 }
 
@@ -448,13 +598,17 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 	BlockNumber insertPage = InvalidBlockNumber;
 	Relation	index = vacuumstate->index;
 	BufferAccessStrategy bas = vacuumstate->bas;
-	bool		useIndexTuple = HnswUseIndexTuple(index);
 
 	/*
-	 * 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);
 
@@ -529,15 +683,9 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 			ntup = (HnswNeighborTuple) PageGetItem(npage, PageGetItemId(npage, neighborOffno));
 
 			/* Overwrite element */
+			/* Use memset instead of MemSet to keep clang-tidy happy */
 			etup->deleted = 1;
-			if (useIndexTuple)
-			{
-				IndexTuple	itup = (IndexTuple) &etup->data;
-
-				MemSet(itup, 0, IndexTupleSize(itup));
-			}
-			else
-				MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
+			memset(&etup->data, 0, VARSIZE_ANY(&etup->data));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)
@@ -608,7 +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);
 }
 
 /*
@@ -617,7 +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);
@@ -635,13 +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

@@ -2,23 +2,33 @@
 
 #include <float.h>
 
+#include "access/genam.h"
+#include "access/generic_xlog.h"
+#include "access/itup.h"
+#include "access/relscan.h"
 #include "access/table.h"
 #include "access/tableam.h"
+#include "access/tupdesc.h"
 #include "access/parallel.h"
 #include "access/xact.h"
-#include "bitvec.h"
+#include "access/xloginsert.h"
 #include "catalog/index.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
 #include "commands/progress.h"
-#include "halfvec.h"
+#include "fmgr.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
+#include "nodes/execnodes.h"
 #include "optimizer/optimizer.h"
 #include "storage/bufmgr.h"
+#include "storage/condition_variable.h"
 #include "tcop/tcopprot.h"
 #include "utils/memutils.h"
-#include "vector.h"
+#include "utils/rel.h"
+#include "utils/sampling.h"
+#include "utils/snapmgr.h"
+#include "utils/tuplesort.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -53,15 +63,13 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/*
-	 * Normalize with KMEANS_NORM_PROC since spherical distance function
-	 * expects unit vectors
+	 * Check with KMEANS_NORM_PROC that the value can be normalized since
+	 * spherical distance function expects unit vectors
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
 		if (!IvfflatCheckNorm(buildstate->kmeansnormprocinfo, buildstate->collation, value))
 			return;
-
-		value = IvfflatNormValue(buildstate->typeInfo, buildstate->collation, value);
 	}
 
 	if (samples->length < targsamples)
@@ -72,7 +80,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	else
 	{
 		if (buildstate->rowstoskip < 0)
-			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, samples->length, targsamples);
+			buildstate->rowstoskip = reservoir_get_next_S(&buildstate->rstate, buildstate->samplerows, targsamples);
 
 		if (buildstate->rowstoskip <= 0)
 		{
@@ -88,6 +96,9 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 
 		buildstate->rowstoskip -= 1;
 	}
+
+	/* Increment after reservoir_get_next_S */
+	buildstate->samplerows += 1;
 }
 
 /*
@@ -124,6 +135,7 @@ SampleRows(IvfflatBuildState * buildstate)
 	int			targsamples = buildstate->samples->maxlen;
 	BlockNumber totalblocks = RelationGetNumberOfBlocks(buildstate->heap);
 
+	buildstate->samplerows = 0;
 	buildstate->rowstoskip = -1;
 
 	BlockSampler_Init(&buildstate->bs, totalblocks, targsamples, RandomInt());
@@ -133,9 +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);
 }
 
 /*
@@ -365,10 +382,20 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
 	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
 	TupleDescInitEntry(buildstate->sortdesc, (AttrNumber) 3, "vector", TupleDescAttr(buildstate->tupdesc, 0)->atttypid, -1, 0);
+#if PG_VERSION_NUM >= 190000
+	TupleDescFinalize(buildstate->sortdesc);
+#endif
 
 	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,
@@ -411,22 +438,30 @@ 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->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)
 	{
-		SampleRows(buildstate);
+		IvfflatBench("sample rows", SampleRows(buildstate));
 
 		if (buildstate->samples->length < buildstate->lists)
 		{
@@ -438,7 +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);
@@ -641,7 +676,11 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
 	ivfspool->sortstate = InitBuildSortState(buildstate.sortdesc, sortmem, coordinate);
 	buildstate.sortstate = ivfspool->sortstate;
 	scan = table_beginscan_parallel(ivfspool->heap,
-									ParallelTableScanFromIvfflatShared(ivfshared));
+									ParallelTableScanFromIvfflatShared(ivfshared)
+#if PG_VERSION_NUM >= 190000
+									,SO_NONE
+#endif
+		);
 	reltuples = table_index_build_scan(ivfspool->heap, ivfspool->index, indexInfo,
 									   true, progress, BuildCallback,
 									   (void *) &buildstate, scan);

src/ivfflat.c

@@ -3,14 +3,19 @@
 #include <float.h>
 
 #include "access/amapi.h"
+#include "access/genam.h"
 #include "access/reloptions.h"
 #include "commands/progress.h"
 #include "commands/vacuum.h"
+#include "fmgr.h"
 #include "ivfflat.h"
+#include "nodes/pg_list.h"
 #include "utils/float.h"
 #include "utils/guc.h"
+#include "utils/relcache.h"
 #include "utils/selfuncs.h"
 #include "utils/spccache.h"
+#include "vector.h"
 
 #if PG_VERSION_NUM < 150000
 #define MarkGUCPrefixReserved(x) EmitWarningsOnPlaceholders(x)
@@ -179,6 +184,64 @@ FUNCTION_PREFIX PG_FUNCTION_INFO_V1(ivfflathandler);
 Datum
 ivfflathandler(PG_FUNCTION_ARGS)
 {
+#if PG_VERSION_NUM >= 190000
+	static const IndexAmRoutine amroutine = {
+		.type = T_IndexAmRoutine,
+		.amstrategies = 0,
+		.amsupport = 5,
+		.amoptsprocnum = 0,
+		.amcanorder = false,
+		.amcanorderbyop = true,
+		.amcanhash = false,
+		.amconsistentequality = false,
+		.amconsistentordering = false,
+		.amcanbackward = false,
+		.amcanunique = false,
+		.amcanmulticol = false,
+		.amoptionalkey = true,
+		.amsearcharray = false,
+		.amsearchnulls = false,
+		.amstorage = false,
+		.amclusterable = false,
+		.ampredlocks = false,
+		.amcanparallel = false,
+		.amcanbuildparallel = true,
+		.amcaninclude = false,
+		.amusemaintenanceworkmem = false,
+		.amsummarizing = false,
+		.amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL,
+		.amkeytype = InvalidOid,
+
+		.ambuild = ivfflatbuild,
+		.ambuildempty = ivfflatbuildempty,
+		.aminsert = ivfflatinsert,
+		.aminsertcleanup = NULL,
+		.ambulkdelete = ivfflatbulkdelete,
+		.amvacuumcleanup = ivfflatvacuumcleanup,
+		.amcanreturn = NULL,
+		.amcostestimate = ivfflatcostestimate,
+		.amgettreeheight = NULL,
+		.amoptions = ivfflatoptions,
+		.amproperty = NULL,
+		.ambuildphasename = ivfflatbuildphasename,
+		.amvalidate = ivfflatvalidate,
+		.amadjustmembers = NULL,
+		.ambeginscan = ivfflatbeginscan,
+		.amrescan = ivfflatrescan,
+		.amgettuple = ivfflatgettuple,
+		.amgetbitmap = NULL,
+		.amendscan = ivfflatendscan,
+		.ammarkpos = NULL,
+		.amrestrpos = NULL,
+		.amestimateparallelscan = NULL,
+		.aminitparallelscan = NULL,
+		.amparallelrescan = NULL,
+		.amtranslatestrategy = NULL,
+		.amtranslatecmptype = NULL,
+	};
+
+	PG_RETURN_POINTER(&amroutine);
+#else
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
@@ -252,4 +315,5 @@ ivfflathandler(PG_FUNCTION_ARGS)
 #endif
 
 	PG_RETURN_POINTER(amroutine);
+#endif
 }

src/ivfflat.h

@@ -9,6 +9,7 @@
 #include "lib/pairingheap.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
+#include "storage/condition_variable.h"
 #include "utils/sampling.h"
 #include "utils/tuplesort.h"
 #include "vector.h"
@@ -25,6 +26,10 @@
 #include "portability/instr_time.h"
 #endif
 
+#if PG_VERSION_NUM >= 190000
+typedef Pointer Item;
+#endif
+
 #define IVFFLAT_MAX_DIM 2000
 
 /* Support functions */
@@ -199,6 +204,7 @@ typedef struct IvfflatBuildState
 	VectorArray samples;
 	VectorArray centers;
 	ListInfo   *listInfo;
+	Size		itemsize;
 
 #ifdef IVFFLAT_KMEANS_DEBUG
 	double		inertia;
@@ -209,7 +215,8 @@ typedef struct IvfflatBuildState
 	/* Sampling */
 	BlockSamplerData bs;
 	ReservoirStateData rstate;
-	int			rowstoskip;
+	double		samplerows;
+	double		rowstoskip;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;
@@ -217,6 +224,7 @@ typedef struct IvfflatBuildState
 	TupleTableSlot *slot;
 
 	/* Memory */
+	Size		memoryUsed;
 	MemoryContext tmpCtx;
 
 	/* Parallel builds */
@@ -297,22 +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/ivfinsert.c

@@ -2,11 +2,16 @@
 
 #include <float.h>
 
+#include "access/genam.h"
 #include "access/generic_xlog.h"
+#include "access/itup.h"
+#include "fmgr.h"
 #include "ivfflat.h"
+#include "nodes/execnodes.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
 #include "utils/memutils.h"
+#include "utils/rel.h"
 
 /*
  * Find the list that minimizes the distance function

src/ivfkmeans.c

@@ -1,17 +1,16 @@
 #include "postgres.h"
 
 #include <float.h>
+#include <limits.h>
 #include <math.h>
 
-#include "bitvec.h"
-#include "halfutils.h"
-#include "halfvec.h"
+#include "access/genam.h"
+#include "fmgr.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
-#include "utils/builtins.h"
 #include "utils/datum.h"
 #include "utils/memutils.h"
-#include "vector.h"
+#include "utils/relcache.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
@@ -101,36 +100,50 @@ NormCenters(const IvfflatTypeInfo * typeInfo, Oid collation, VectorArray centers
 	MemoryContext normCtx = AllocSetContextCreate(CurrentMemoryContext,
 												  "Ivfflat norm temporary context",
 												  ALLOCSET_DEFAULT_SIZES);
-	MemoryContext oldCtx = MemoryContextSwitchTo(normCtx);
 
-	for (int j = 0; j < centers->length; j++)
-	{
-		Datum		center = PointerGetDatum(VectorArrayGet(centers, j));
-		Datum		newCenter = IvfflatNormValue(typeInfo, collation, center);
-		Size		size = VARSIZE_ANY(DatumGetPointer(newCenter));
-
-		if (size > centers->itemsize)
-			elog(ERROR, "safety check failed");
-
-		memcpy(DatumGetPointer(center), DatumGetPointer(newCenter), size);
-		MemoryContextReset(normCtx);
-	}
-
-	MemoryContextSwitchTo(oldCtx);
+	IvfflatNormVectors(typeInfo, collation, centers, normCtx);
 	MemoryContextDelete(normCtx);
 }
 
 /*
- * Quick approach if we have no data
+ * Check if vector array contains a vector
+ */
+static bool
+VectorArrayContains(VectorArray arr, Pointer val)
+{
+	Datum		d = PointerGetDatum(val);
+
+	for (int i = 0; i < arr->length; i++)
+	{
+		if (datumIsEqual(d, PointerGetDatum(VectorArrayGet(arr, i)), false, -1))
+			return true;
+	}
+	return false;
+}
+
+/*
+ * Quick approach if we have little data
  */
 static void
-RandomCenters(Relation index, VectorArray centers, const IvfflatTypeInfo * typeInfo)
+QuickCenters(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo)
 {
 	int			dimensions = centers->dim;
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 	Oid			collation = index->rd_indcollation[0];
 	float	   *x = (float *) palloc(sizeof(float) * dimensions);
 
+	/* Fill with unique samples (already normalized) */
+	for (int i = 0; i < samples->length; i++)
+	{
+		Pointer		sample = VectorArrayGet(samples, i);
+
+		if (!VectorArrayContains(centers, sample))
+		{
+			VectorArraySet(centers, centers->length, sample);
+			centers->length++;
+		}
+	}
+
 	/* Fill with random data */
 	while (centers->length < centers->maxlen)
 	{
@@ -260,7 +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;
@@ -279,8 +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;
@@ -292,18 +303,13 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const Ivff
 	Size		newcdistSize = sizeof(float) * numCenters;
 
 	/* Calculate total size */
-	Size		totalSize = samplesSize + centersSize + newCentersSize + aggSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
+	Size		totalSize = memoryUsed + newCentersSize + aggSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
 
 	/* Check memory requirements */
-	/* Add one to error message to ceil */
-	if (totalSize > (Size) maintenance_work_mem * 1024L)
-		ereport(ERROR,
-				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
-				 errmsg("memory required is %zu MB, maintenance_work_mem is %d MB",
-						totalSize / (1024 * 1024) + 1, maintenance_work_mem / 1024)));
+	IvfflatCheckMemoryUsage(totalSize);
 
 	/* Ensure indexing does not overflow */
-	if (numCenters * numCenters > INT_MAX)
+	if (numCenters > INT_MAX / numCenters)
 		elog(ERROR, "Indexing overflow detected. Please report a bug.");
 
 	/* Set support functions */
@@ -564,17 +570,17 @@ CheckCenters(Relation index, VectorArray centers, const IvfflatTypeInfo * typeIn
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo)
+IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, const IvfflatTypeInfo * typeInfo, Size memoryUsed)
 {
 	MemoryContext kmeansCtx = AllocSetContextCreate(CurrentMemoryContext,
 													"Ivfflat kmeans temporary context",
 													ALLOCSET_DEFAULT_SIZES);
 	MemoryContext oldCtx = MemoryContextSwitchTo(kmeansCtx);
 
-	if (samples->length == 0)
-		RandomCenters(index, centers, typeInfo);
+	if (samples->length <= centers->maxlen)
+		QuickCenters(index, samples, centers, typeInfo);
 	else
-		ElkanKmeans(index, samples, centers, typeInfo);
+		ElkanKmeans(index, samples, centers, typeInfo, memoryUsed);
 
 	CheckCenters(index, centers, typeInfo);
 

src/ivfscan.c

@@ -2,15 +2,26 @@
 
 #include <float.h>
 
+#include "access/genam.h"
+#include "access/itup.h"
 #include "access/relscan.h"
+#include "access/tupdesc.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
+#include "fmgr.h"
 #include "lib/pairingheap.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
 #include "pgstat.h"
 #include "storage/bufmgr.h"
 #include "utils/memutils.h"
+#include "utils/rel.h"
+#include "utils/snapmgr.h"
+#include "utils/tuplesort.h"
+
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
 
 #define GetScanList(ptr) pairingheap_container(IvfflatScanList, ph_node, ptr)
 #define GetScanListConst(ptr) pairingheap_const_container(IvfflatScanList, ph_node, ptr)
@@ -287,6 +298,9 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	so->tupdesc = CreateTemplateTupleDesc(2);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
+#if PG_VERSION_NUM >= 190000
+	TupleDescFinalize(so->tupdesc);
+#endif
 
 	/* Prep sort */
 	so->sortstate = InitScanSortState(so->tupdesc);

src/ivfutils.c

@@ -1,13 +1,21 @@
 #include "postgres.h"
 
+#include "access/genam.h"
 #include "access/generic_xlog.h"
-#include "bitvec.h"
-#include "catalog/pg_type.h"
 #include "fmgr.h"
 #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"
+
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
 
 /*
  * Allocate a vector array
@@ -15,11 +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;
@@ -82,6 +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/ivfvacuum.c

@@ -1,9 +1,12 @@
 #include "postgres.h"
 
+#include "access/genam.h"
 #include "access/generic_xlog.h"
+#include "access/itup.h"
 #include "commands/vacuum.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "utils/relcache.h"
 
 #if PG_VERSION_NUM >= 180000
 #define vacuum_delay_point() vacuum_delay_point(false)

src/sparsevec.c

@@ -5,18 +5,27 @@
 
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
-#include "common/string.h"
 #include "fmgr.h"
 #include "halfutils.h"
 #include "halfvec.h"
+#include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
 #include "sparsevec.h"
 #include "utils/array.h"
 #include "utils/builtins.h"
 #include "utils/float.h"
+#include "utils/fmgrprotos.h"
 #include "utils/lsyscache.h"
 #include "vector.h"
 
+#if PG_VERSION_NUM >= 160000
+#include "varatt.h"
+#endif
+
+#if PG_VERSION_NUM >= 170000
+#include "parser/scansup.h"
+#endif
+
 typedef struct SparseInputElement
 {
 	int32		index;
@@ -150,9 +159,9 @@ InitSparseVector(int dim, int nnz)
 	return result;
 }
 
-/*
- * Check for whitespace, since array_isspace() is static
- */
+#if PG_VERSION_NUM >= 170000
+#define sparsevec_isspace(ch) scanner_isspace(ch)
+#else
 static inline bool
 sparsevec_isspace(char ch)
 {
@@ -165,6 +174,7 @@ sparsevec_isspace(char ch)
 		return true;
 	return false;
 }
+#endif
 
 /*
  * Compare indices
@@ -172,10 +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

@@ -16,16 +16,20 @@
 #include "port.h"				/* for strtof() */
 #include "sparsevec.h"
 #include "utils/array.h"
-#include "utils/builtins.h"
 #include "utils/float.h"
+#include "utils/fmgrprotos.h"
 #include "utils/lsyscache.h"
-#include "utils/numeric.h"
+#include "utils/varbit.h"
 #include "vector.h"
 
 #if PG_VERSION_NUM >= 160000
 #include "varatt.h"
 #endif
 
+#if PG_VERSION_NUM >= 170000
+#include "parser/scansup.h"
+#endif
+
 #define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
 #define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
 
@@ -36,7 +40,7 @@
 #endif
 
 #if PG_VERSION_NUM >= 180000
-PG_MODULE_MAGIC_EXT(.name = "vector",.version = "0.8.1");
+PG_MODULE_MAGIC_EXT(.name = "vector", .version = "0.8.5");
 #else
 PG_MODULE_MAGIC;
 #endif
@@ -129,9 +133,9 @@ InitVector(int dim)
 	return result;
 }
 
-/*
- * Check for whitespace, since array_isspace() is static
- */
+#if PG_VERSION_NUM >= 170000
+#define vector_isspace(ch) scanner_isspace(ch)
+#else
 static inline bool
 vector_isspace(char ch)
 {
@@ -144,6 +148,7 @@ vector_isspace(char ch)
 		return true;
 	return false;
 }
+#endif
 
 /*
  * Check state array

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/005_ivfflat_query_recall.pl

@@ -16,29 +16,19 @@ $node->safe_psql("postgres",
 	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
 );
 
-# Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+# Add index
+$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
 
-for my $i (0 .. $#operators)
+# Test 100% recall
+for (1 .. 20)
 {
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
-
-	# Add index
-	$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v $opclass);");
-
-	# Test 100% recall
-	for (1 .. 20)
-	{
-		my $id = int(rand() * 100000);
-		my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $id;");
-		my $res = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SELECT v FROM tst ORDER BY v <-> '$query' LIMIT 1;
-		));
-		is($res, $query);
-	}
+	my $id = int(rand() * 100000);
+	my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $id;");
+	my $res = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		SELECT v FROM tst ORDER BY v <-> '$query' LIMIT 1;
+	));
+	is($res, $query);
 }
 
 done_testing();

test/t/045_hnsw_hqann.pl

@@ -1,113 +0,0 @@
-use strict;
-use warnings FATAL => 'all';
-use PostgreSQL::Test::Cluster;
-use PostgreSQL::Test::Utils;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @cs = ();
-my @expected;
-my $limit = 20;
-my $dim = 3;
-my $array_sql = join(",", ('random()') x $dim);
-my $nc = 1000;
-
-sub test_recall
-{
-	my ($min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		EXPLAIN ANALYZE SELECT i FROM tst WHERE c = $cs[0] ORDER BY v $operator '$queries[0]' LIMIT $limit;
-	));
-	like($explain, qr/Index Cond/);
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SELECT i FROM tst WHERE c = $cs[$i] ORDER BY v $operator '$queries[$i]' LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-		my %actual_set = map { $_ => 1 } @actual_ids;
-
-		is(scalar(@actual_ids), $limit);
-
-		my @expected_ids = split("\n", $expected[$i]);
-
-		foreach (@expected_ids)
-		{
-			if (exists($actual_set{$_}))
-			{
-				$correct++;
-			}
-			$total++;
-		}
-	}
-
-	cmp_ok($correct / $total, ">=", $min, $operator);
-}
-
-# Initialize node
-$node = PostgreSQL::Test::Cluster->new('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim), c int4);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql], i % $nc FROM generate_series(1, 20000) i;"
-);
-
-# Generate queries
-for (1 .. 20)
-{
-	my @r = ();
-	for (1 .. $dim)
-	{
-		push(@r, rand());
-	}
-	push(@queries, "[" . join(",", @r) . "]");
-	push(@cs, int(rand() * $nc));
-}
-
-# Get exact results
-@expected = ();
-for my $i (0 .. $#queries)
-{
-	my $res = $node->safe_psql("postgres", "SELECT i FROM tst WHERE c = $cs[$i] ORDER BY v <=> '$queries[$i]' LIMIT $limit;");
-	push(@expected, $res);
-}
-
-# Add index
-$node->safe_psql("postgres", qq(
-	SET maintenance_work_mem = '256MB';
-	SET max_parallel_maintenance_workers = 2;
-	CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, c);
-));
-
-# Test recall
-test_recall(0.99, '<=>');
-
-# Test vacuum
-$node->safe_psql("postgres", "DELETE FROM tst WHERE c > 5;");
-$node->safe_psql("postgres", "VACUUM tst;");
-
-# Test columns
-my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c, v vector_cosine_ops);");
-like($stderr, qr/first column must be a vector/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, c, c);");
-like($stderr, qr/index cannot have more than two columns/);
-
-($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_cosine_ops, v vector_cosine_ops);");
-like($stderr, qr/column 2 cannot be a vector/);
-
-done_testing();

test/t/045_hnsw_low_memory_build.pl

@@ -0,0 +1,29 @@
+use strict;
+use warnings FATAL => 'all';
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+# Initialize node
+my $node = PostgreSQL::Test::Cluster->new('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (v vector(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 1000) i;"
+);
+
+my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+	SET client_min_messages = DEBUG;
+	SET maintenance_work_mem = '3073kB';
+	ALTER TABLE tst SET (parallel_workers = 1);
+	CREATE INDEX ON tst USING hnsw (v vector_l2_ops);
+));
+is($ret, 0, $stderr);
+like($stderr, qr/using \d+ parallel workers/);
+like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem after 0 tuples/);
+
+done_testing();

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 = '0.8.1'
+default_version = '0.8.5'
 module_pathname = '$libdir/vector'
 relocatable = true