Include vectors with zero norm in indexes for cosine distance

The HnswInitElement() allocate an element with not initialize value
filed, which may has garbage that lead HnswFreeElement() free an
invalid pointer.

.github/workflows/build.yml

@@ -8,6 +8,8 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - postgres: 17
+            os: ubuntu-22.04
           - postgres: 16
             os: ubuntu-22.04
           - postgres: 15
@@ -21,7 +23,7 @@ jobs:
           - postgres: 11
             os: ubuntu-20.04
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: ${{ matrix.postgres }}
@@ -43,7 +45,7 @@ jobs:
     runs-on: macos-latest
     if: ${{ !startsWith(github.ref_name, 'windows') }}
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: 14
@@ -65,7 +67,7 @@ jobs:
     runs-on: windows-latest
     if: ${{ !startsWith(github.ref_name, 'mac') }}
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: ankane/setup-postgres@v1
         with:
           postgres-version: 14

CHANGELOG.md

@@ -1,7 +1,11 @@
-## 0.5.1 (unreleased)
+## 0.5.2 (unreleased)
 
-- Added check for MVCC-compliant snapshot for HNSW index scans
+- Added support for on-disk parallel index builds for HNSW
-- Improved performance of index scans for IVFFlat after updates and deletes
+
+## 0.5.1 (2023-10-10)
+
+- Improved performance of HNSW index builds
+- Added check for MVCC-compliant snapshot for index scans
 
 ## 0.5.0 (2023-08-28)
 

META.json

@@ -2,7 +2,7 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.5.0",
+	"version": "0.5.1",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.5.0",
+			"version": "0.5.1",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,10 +1,10 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+EXTVERSION = 0.5.1
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/tinyint.o src/vector.o
+OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
-HEADERS = src/tinyint.h src/vector.h
+HEADERS = src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))

Makefile.win

@@ -1,8 +1,8 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+EXTVERSION = 0.5.1
 
-OBJS = 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\tinyint.obj src\vector.obj
+OBJS = 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\vector.obj
-HEADERS = src\tinyint.h src\vector.h
+HEADERS = src\vector.h
 
 REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
 REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)

README.md

@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
 
 ```sh
 cd /tmp
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -26,7 +26,7 @@ make install # may need sudo
 
 See the [installation notes](#installation-notes) if you run into issues
 
-You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres)
+You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
 
 ## Getting Started
 
@@ -215,6 +215,23 @@ SELECT ...
 COMMIT;
 ```
 
+### Indexing Progress
+
+Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
+
+```sql
+SELECT phase, round(100.0 * tuples_done / nullif(tuples_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
+```
+
+The phases for IVFFlat are:
+
+1. `initializing`
+2. `performing k-means`
+3. `assigning tuples`
+4. `loading tuples`
+
+Note: `%` is only populated during the `loading tuples` phase
+
 ## HNSW
 
 An HNSW index creates a multilayer graph. It has slower build times and uses more memory than IVFFlat, but has better query performance (in terms of speed-recall tradeoff). There’s no training step like IVFFlat, so the index can be created without any data in the table.
@@ -252,6 +269,8 @@ Specify HNSW parameters
 CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WITH (m = 16, ef_construction = 64);
 ```
 
+A higher value of `ef_construction` provides better recall at the cost of index build time / insert speed.
+
 ### Query Options
 
 Specify the size of the dynamic candidate list for search (40 by default)
@@ -271,22 +290,18 @@ SELECT ...
 COMMIT;
 ```
 
-## Indexing Progress
+### Indexing Progress
 
 Check [indexing progress](https://www.postgresql.org/docs/current/progress-reporting.html#CREATE-INDEX-PROGRESS-REPORTING) with Postgres 12+
 
 ```sql
-SELECT phase, tuples_done, tuples_total FROM pg_stat_progress_create_index;
+SELECT phase, round(100.0 * blocks_done / nullif(blocks_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
 ```
 
-The phases are:
+The phases for HNSW are:
 
 1. `initializing`
-2. `performing k-means` - IVFFlat only
+2. `loading tuples`
-3. `assigning tuples` - IVFFlat only
-4. `loading tuples`
-
-Note: `tuples_done` and `tuples_total` are only populated during the `loading tuples` phase
 
 ## Filtering
 
@@ -317,13 +332,15 @@ CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(cate
 
 ## Hybrid Search
 
-Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search ([Python example](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py)).
+Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search.
 
 ```sql
 SELECT id, content FROM items, plainto_tsquery('hello search') query
     WHERE textsearch @@ query ORDER BY ts_rank_cd(textsearch, query) DESC LIMIT 5;
 ```
 
+You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
+
 ## Performance
 
 Use `EXPLAIN ANALYZE` to debug performance.
@@ -360,6 +377,7 @@ Use pgvector from any language with a Postgres client. You can even generate and
 
 Language | Libraries / Examples
 --- | ---
+C | [pgvector-c](https://github.com/pgvector/pgvector-c)
 C++ | [pgvector-cpp](https://github.com/pgvector/pgvector-cpp)
 C# | [pgvector-dotnet](https://github.com/pgvector/pgvector-dotnet)
 Crystal | [pgvector-crystal](https://github.com/pgvector/pgvector-crystal)
@@ -367,10 +385,11 @@ Dart | [pgvector-dart](https://github.com/pgvector/pgvector-dart)
 Elixir | [pgvector-elixir](https://github.com/pgvector/pgvector-elixir)
 Go | [pgvector-go](https://github.com/pgvector/pgvector-go)
 Haskell | [pgvector-haskell](https://github.com/pgvector/pgvector-haskell)
-Java, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
+Java, Kotlin, Groovy, Scala | [pgvector-java](https://github.com/pgvector/pgvector-java)
+JavaScript, TypeScript | [pgvector-node](https://github.com/pgvector/pgvector-node)
 Julia | [pgvector-julia](https://github.com/pgvector/pgvector-julia)
 Lua | [pgvector-lua](https://github.com/pgvector/pgvector-lua)
-Node.js | [pgvector-node](https://github.com/pgvector/pgvector-node)
+Nim | [pgvector-nim](https://github.com/pgvector/pgvector-nim)
 Perl | [pgvector-perl](https://github.com/pgvector/pgvector-perl)
 PHP | [pgvector-php](https://github.com/pgvector/pgvector-php)
 Python | [pgvector-python](https://github.com/pgvector/pgvector-python)
@@ -378,6 +397,7 @@ R | [pgvector-r](https://github.com/pgvector/pgvector-r)
 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)
+Zig | [pgvector-zig](https://github.com/pgvector/pgvector-zig)
 
 ## Frequently Asked Questions
 
@@ -393,6 +413,55 @@ Yes, pgvector uses the write-ahead log (WAL), which allows for replication and p
 
 You’ll need to use [dimensionality reduction](https://en.wikipedia.org/wiki/Dimensionality_reduction) at the moment.
 
+#### Can I store vectors with different dimensions in the same column?
+
+You can use `vector` as the type (instead of `vector(3)`).
+
+```sql
+CREATE TABLE embeddings (model_id bigint, item_id bigint, embedding vector, PRIMARY KEY (model_id, item_id));
+```
+
+However, you can only create indexes on rows with the same number of dimensions (using [expression](https://www.postgresql.org/docs/current/indexes-expressional.html) and [partial](https://www.postgresql.org/docs/current/indexes-partial.html) indexing):
+
+```sql
+CREATE INDEX ON embeddings USING hnsw ((embedding::vector(3)) vector_l2_ops) WHERE (model_id = 123);
+```
+
+and query with:
+
+```sql
+SELECT * FROM embeddings WHERE model_id = 123 ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
+```
+
+#### Can I store vectors with more precision?
+
+You can use the `double precision[]` or `numeric[]` type to store vectors with more precision.
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding double precision[]);
+
+-- use {} instead of [] for Postgres arrays
+INSERT INTO items (embedding) VALUES ('{1,2,3}'), ('{4,5,6}');
+```
+
+Optionally, add a [check constraint](https://www.postgresql.org/docs/current/ddl-constraints.html) to ensure data can be converted to the `vector` type and has the expected dimensions.
+
+```sql
+ALTER TABLE items ADD CHECK (vector_dims(embedding::vector) = 3);
+```
+
+Use [expression indexing](https://www.postgresql.org/docs/current/indexes-expressional.html) to index (at a lower precision):
+
+```sql
+CREATE INDEX ON items USING hnsw ((embedding::vector(3)) vector_l2_ops);
+```
+
+and query with:
+
+```sql
+SELECT * FROM items ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
+```
+
 ## Troubleshooting
 
 #### Why isn’t a query using an index?
@@ -406,6 +475,8 @@ SELECT ...
 COMMIT;
 ```
 
+Also, if the table is small, a table scan may be faster.
+
 #### Why isn’t a query using a parallel table scan?
 
 The planner doesn’t consider [out-of-line storage](https://www.postgresql.org/docs/current/storage-toast.html) in cost estimates, which can make a serial scan look cheaper. You can reduce the cost of a parallel scan for a query with:
@@ -509,7 +580,7 @@ Then use `nmake` to build:
 
 ```cmd
 set "PGROOT=C:\Program Files\PostgreSQL\15"
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -530,7 +601,7 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
 ```
@@ -595,7 +666,7 @@ pgvector is available on [these providers](https://github.com/pgvector/pgvector/
 
 ## Upgrading
 
-Install the latest version. Then in each database you want to upgrade, run:
+[Install](#installation) the latest version (use the same method as the original installation). Then in each database you want to upgrade, run:
 
 ```sql
 ALTER EXTENSION vector UPDATE;

sql/vector--0.5.0--0.5.1.sql

@@ -1,67 +1,2 @@
 -- complain if script is sourced in psql, rather than via CREATE EXTENSION
 \echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit
-
--- tinyint
-
-CREATE TYPE tinyint;
-
-CREATE FUNCTION tinyint_in(cstring, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_out(tinyint) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_recv(internal, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_send(tinyint) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE tinyint (
-	INPUT     = tinyint_in,
-	OUTPUT    = tinyint_out,
-	RECEIVE   = tinyint_recv,
-	SEND      = tinyint_send,
-	INTERNALLENGTH = 1,
-	PASSEDBYVALUE,
-	ALIGNMENT = char
-);
-
-CREATE FUNCTION integer_to_tinyint(integer, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION numeric_to_tinyint(numeric, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (integer AS tinyint)
-	WITH FUNCTION integer_to_tinyint(integer, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (numeric AS tinyint)
-	WITH FUNCTION numeric_to_tinyint(numeric, integer, boolean) AS IMPLICIT;
-
-CREATE FUNCTION l2_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_negative_inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR <-> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = tinyint_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

sql/vector.sql

@@ -290,68 +290,3 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
-
--- tinyint
-
-CREATE TYPE tinyint;
-
-CREATE FUNCTION tinyint_in(cstring, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_out(tinyint) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_recv(internal, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_send(tinyint) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE tinyint (
-	INPUT     = tinyint_in,
-	OUTPUT    = tinyint_out,
-	RECEIVE   = tinyint_recv,
-	SEND      = tinyint_send,
-	INTERNALLENGTH = 1,
-	PASSEDBYVALUE,
-	ALIGNMENT = char
-);
-
-CREATE FUNCTION integer_to_tinyint(integer, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION numeric_to_tinyint(numeric, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (integer AS tinyint)
-	WITH FUNCTION integer_to_tinyint(integer, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (numeric AS tinyint)
-	WITH FUNCTION numeric_to_tinyint(numeric, integer, boolean) AS IMPLICIT;
-
-CREATE FUNCTION l2_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_negative_inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR <-> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = tinyint_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

src/hnsw.c

@@ -14,6 +14,7 @@
 #endif
 
 int			hnsw_ef_search;
+bool		hnsw_enable_parallel_build;
 static relopt_kind hnsw_relopt_kind;
 
 /*
@@ -39,6 +40,11 @@ HnswInit(void)
 	DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
 							"Valid range is 1..1000.", &hnsw_ef_search,
 							HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
+
+	/* Behind a variable for now since can be slower than building in memory */
+	DefineCustomBoolVariable("hnsw.enable_parallel_build", "Enables or disables building indexes in parallel",
+							 NULL, &hnsw_enable_parallel_build,
+							 false, PGC_USERSET, 0, NULL, NULL, NULL);
 }
 
 /*

src/hnsw.h

@@ -4,6 +4,7 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
+#include "access/parallel.h"
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
@@ -14,6 +15,10 @@
 #error "Requires PostgreSQL 11+"
 #endif
 
+#if PG_VERSION_NUM < 120000
+#include "access/relscan.h"
+#endif
+
 #define HNSW_MAX_DIM 2000
 
 /* Support functions */
@@ -59,7 +64,7 @@
 
 #define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
 
-#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
+#define HNSW_ELEMENT_TUPLE_SIZE(size)	MAXALIGN(offsetof(HnswElementTupleData, data) + (size))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -90,6 +95,7 @@
 
 /* Variables */
 extern int	hnsw_ef_search;
+extern bool hnsw_enable_parallel_build;
 
 typedef struct HnswNeighborArray HnswNeighborArray;
 
@@ -103,7 +109,7 @@ typedef struct HnswElementData
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Vector	   *vec;
+	Datum		value;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -112,11 +118,13 @@ typedef struct HnswCandidate
 {
 	HnswElement element;
 	float		distance;
+	bool		closer;
 }			HnswCandidate;
 
 typedef struct HnswNeighborArray
 {
 	int			length;
+	bool		closerSet;
 	HnswCandidate *items;
 }			HnswNeighborArray;
 
@@ -134,6 +142,49 @@ typedef struct HnswOptions
 	int			efConstruction; /* size of dynamic candidate list */
 }			HnswOptions;
 
+typedef struct HnswSpool
+{
+	Relation	heap;
+	Relation	index;
+}			HnswSpool;
+
+typedef struct HnswShared
+{
+	/* Immutable state */
+	Oid			heaprelid;
+	Oid			indexrelid;
+	bool		isconcurrent;
+	int			scantuplesortstates;
+
+	/* Worker progress */
+	ConditionVariable workersdonecv;
+
+	/* Mutex for mutable state */
+	slock_t		mutex;
+
+	/* Mutable state */
+	int			nparticipantsdone;
+	double		reltuples;
+	double		indtuples;
+
+#if PG_VERSION_NUM < 120000
+	ParallelHeapScanDescData heapdesc;	/* must come last */
+#endif
+}			HnswShared;
+
+#if PG_VERSION_NUM >= 120000
+#define ParallelTableScanFromHnswShared(shared) \
+	(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(HnswShared)))
+#endif
+
+typedef struct HnswLeader
+{
+	ParallelContext *pcxt;
+	int			nparticipanttuplesorts;
+	HnswShared *hnswshared;
+	Snapshot	snapshot;
+}			HnswLeader;
+
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -161,12 +212,16 @@ typedef struct HnswBuildState
 	HnswElement entryPoint;
 	double		ml;
 	int			maxLevel;
-	double		maxInMemoryElements;
+	long		memoryLeft;
 	bool		flushed;
 	Vector	   *normvec;
 
 	/* Memory */
 	MemoryContext tmpCtx;
+
+	/* Parallel builds */
+	HnswLeader *hnswleader;
+	HnswShared *hnswshared;
 }			HnswBuildState;
 
 typedef struct HnswMetaPageData
@@ -202,7 +257,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		vec;
+	Vector		data;
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -261,7 +316,7 @@ typedef struct HnswVacuumState
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
-bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
+void		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 void		HnswCommitBuffer(Buffer buf, GenericXLogState *state);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
@@ -287,6 +342,7 @@ void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation i
 void		HnswSetElementTuple(HnswElementTuple etup, HnswElement element);
 void		HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
 void		HnswLoadNeighbors(HnswElement element, Relation index, int m);
+PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
 
 /* Index access methods */
 IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);

src/hnswbuild.c

@@ -2,12 +2,16 @@
 
 #include <math.h>
 
+#include "access/parallel.h"
+#include "access/xact.h"
 #include "catalog/index.h"
 #include "hnsw.h"
 #include "miscadmin.h"
 #include "lib/pairingheap.h"
 #include "nodes/pg_list.h"
 #include "storage/bufmgr.h"
+#include "tcop/tcopprot.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -35,6 +39,23 @@
 #define UpdateProgress(index, val) ((void)val)
 #endif
 
+#if PG_VERSION_NUM >= 140000
+#include "utils/backend_status.h"
+#include "utils/wait_event.h"
+#endif
+
+#if PG_VERSION_NUM >= 120000
+#include "access/table.h"
+#include "optimizer/optimizer.h"
+#else
+#include "access/heapam.h"
+#include "optimizer/planner.h"
+#include "pgstat.h"
+#endif
+
+#define PARALLEL_KEY_HNSW_SHARED		UINT64CONST(0xA000000000000001)
+#define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000002)
+
 /*
  * Create the metapage
  */
@@ -105,8 +126,7 @@ CreateElementPages(HnswBuildState * buildstate)
 {
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
-	int			dimensions = buildstate->dimensions;
+	Size		etupAllocSize;
-	Size		etupSize;
 	Size		maxSize;
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
@@ -117,11 +137,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	ListCell   *lc;
 
 	/* Calculate sizes */
+	etupAllocSize = BLCKSZ;
 	maxSize = HNSW_MAX_SIZE;
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
-	etup = palloc0(etupSize);
+	etup = palloc0(etupAllocSize);
 	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
@@ -133,15 +153,24 @@ CreateElementPages(HnswBuildState * buildstate)
 	foreach(lc, buildstate->elements)
 	{
 		HnswElement element = lfirst(lc);
+		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
 
-		HnswSetElementTuple(etup, element);
+		/* Zero memory for each element */
+		MemSet(etup, 0, etupAllocSize);
 
 		/* Calculate sizes */
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(element->value)));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
+		/* Initial size check */
+		if (etupSize > etupAllocSize)
+			elog(ERROR, "index tuple too large");
+
+		HnswSetElementTuple(etup, element);
+
 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
 			HnswBuildAppendPage(index, &buf, &page, &state, forkNum);
@@ -264,26 +293,26 @@ FlushPages(HnswBuildState * buildstate)
  * Insert tuple
  */
 static bool
-InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup)
+InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup, MemoryContext outerCtx)
 {
 	FmgrInfo   *procinfo = buildstate->procinfo;
 	Oid			collation = buildstate->collation;
 	HnswElement entryPoint = buildstate->entryPoint;
 	int			efConstruction = buildstate->efConstruction;
 	int			m = buildstate->m;
+	MemoryContext oldCtx;
 
 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
-	{
+		HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec);
-		if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
-			return false;
-	}
 
 	/* Copy value to element so accessible outside of memory context */
-	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
+	oldCtx = MemoryContextSwitchTo(outerCtx);
+	element->value = datumCopy(value, false, -1);
+	MemoryContextSwitchTo(oldCtx);
 
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
@@ -313,6 +342,21 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	return *dup == NULL;
 }
 
+/*
+ * Get the memory used by an element
+ */
+static long
+HnswElementMemory(HnswElement e, int m)
+{
+	long		elementSize = sizeof(HnswElementData);
+
+	elementSize += sizeof(HnswNeighborArray) * (e->level + 1);
+	elementSize += sizeof(HnswCandidate) * (m * (e->level + 2));
+	elementSize += sizeof(ItemPointerData);
+	elementSize += VARSIZE_ANY(DatumGetPointer(e->value));
+	return elementSize;
+}
+
 /*
  * Callback for table_index_build_scan
  */
@@ -334,7 +378,7 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (isnull[0])
 		return;
 
-	if (buildstate->indtuples >= buildstate->maxInMemoryElements)
+	if (buildstate->memoryLeft <= 0)
 	{
 		if (!buildstate->flushed)
 		{
@@ -349,7 +393,18 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap))
-			UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+		{
+			if (buildstate->hnswshared)
+			{
+				HnswShared *hnswshared = buildstate->hnswshared;
+
+				SpinLockAcquire(&hnswshared->mutex);
+				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++hnswshared->indtuples);
+				SpinLockRelease(&hnswshared->mutex);
+			}
+			else
+				UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+		}
 
 		/* Reset memory context */
 		MemoryContextSwitchTo(oldCtx);
@@ -360,13 +415,12 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Allocate necessary memory outside of memory context */
 	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
-	element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
 
 	/* Use memory context since detoast can allocate */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
 
 	/* Insert tuple */
-	inserted = InsertTuple(index, values, element, buildstate, &dup);
+	inserted = InsertTuple(index, values, element, buildstate, &dup, oldCtx);
 
 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
@@ -374,31 +428,21 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Add outside memory context */
 	if (dup != NULL)
+	{
 		HnswAddHeapTid(dup, tid);
+		buildstate->memoryLeft -= sizeof(ItemPointerData);
+	}
 
 	/* Add to buildstate or free */
 	if (inserted)
+	{
 		buildstate->elements = lappend(buildstate->elements, element);
+		buildstate->memoryLeft -= HnswElementMemory(element, buildstate->m);
+	}
 	else
 		HnswFreeElement(element);
 }
 
-/*
- * Get the max number of elements that fit into maintenance_work_mem
- */
-static double
-HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
-{
-	Size		elementSize = sizeof(HnswElementData);
-	double		avgLevel = -log(0.5) * ml;
-
-	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
-	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
-	elementSize += sizeof(ItemPointerData);
-	elementSize += VECTOR_SIZE(dimensions);
-	return (maintenance_work_mem * 1024L) / elementSize;
-}
-
 /*
  * Initialize the build state
  */
@@ -436,7 +480,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->entryPoint = NULL;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->maxInMemoryElements = HnswGetMaxInMemoryElements(buildstate->m, buildstate->ml, buildstate->dimensions);
+	buildstate->memoryLeft = maintenance_work_mem * 1024L;
 	buildstate->flushed = false;
 
 	/* Reuse for each tuple */
@@ -445,6 +489,9 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);
+
+	buildstate->hnswleader = NULL;
+	buildstate->hnswshared = NULL;
 }
 
 /*
@@ -457,21 +504,374 @@ FreeBuildState(HnswBuildState * buildstate)
 	MemoryContextDelete(buildstate->tmpCtx);
 }
 
+/*
+ * Within leader, wait for end of heap scan
+ */
+static double
+ParallelHeapScan(HnswBuildState * buildstate)
+{
+	HnswShared *hnswshared = buildstate->hnswleader->hnswshared;
+	int			nparticipanttuplesorts;
+	double		reltuples;
+
+	nparticipanttuplesorts = buildstate->hnswleader->nparticipanttuplesorts;
+	for (;;)
+	{
+		SpinLockAcquire(&hnswshared->mutex);
+		if (hnswshared->nparticipantsdone == nparticipanttuplesorts)
+		{
+			buildstate->indtuples = hnswshared->indtuples;
+			reltuples = hnswshared->reltuples;
+			SpinLockRelease(&hnswshared->mutex);
+			break;
+		}
+		SpinLockRelease(&hnswshared->mutex);
+
+		ConditionVariableSleep(&hnswshared->workersdonecv,
+							   WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
+	}
+
+	ConditionVariableCancelSleep();
+
+	return reltuples;
+}
+
+/*
+ * Perform a worker's portion of a parallel insert
+ */
+static void
+HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool progress)
+{
+	HnswBuildState buildstate;
+#if PG_VERSION_NUM >= 120000
+	TableScanDesc scan;
+#else
+	HeapScanDesc scan;
+#endif
+	double		reltuples;
+	IndexInfo  *indexInfo;
+
+	/* Join parallel scan */
+	indexInfo = BuildIndexInfo(hnswspool->index);
+	indexInfo->ii_Concurrent = hnswshared->isconcurrent;
+	InitBuildState(&buildstate, hnswspool->heap, hnswspool->index, indexInfo, MAIN_FORKNUM);
+	/* TODO Support in-memory builds */
+	buildstate.memoryLeft = 0;
+	buildstate.flushed = true;
+	buildstate.hnswshared = hnswshared;
+#if PG_VERSION_NUM >= 120000
+	scan = table_beginscan_parallel(hnswspool->heap,
+									ParallelTableScanFromHnswShared(hnswshared));
+	reltuples = table_index_build_scan(hnswspool->heap, hnswspool->index, indexInfo,
+									   true, progress, BuildCallback,
+									   (void *) &buildstate, scan);
+#else
+	scan = heap_beginscan_parallel(hnswspool->heap, &hnswshared->heapdesc);
+	reltuples = IndexBuildHeapScan(hnswspool->heap, hnswspool->index, indexInfo,
+								   true, BuildCallback,
+								   (void *) &buildstate, scan);
+#endif
+
+	/* Record statistics */
+	SpinLockAcquire(&hnswshared->mutex);
+	hnswshared->nparticipantsdone++;
+	hnswshared->reltuples += reltuples;
+	SpinLockRelease(&hnswshared->mutex);
+
+	/* Log statistics */
+	if (progress)
+		ereport(DEBUG1, (errmsg("leader processed " INT64_FORMAT " tuples", (int64) reltuples)));
+	else
+		ereport(DEBUG1, (errmsg("worker processed " INT64_FORMAT " tuples", (int64) reltuples)));
+
+	/* Notify leader */
+	ConditionVariableSignal(&hnswshared->workersdonecv);
+
+	FreeBuildState(&buildstate);
+}
+
+/*
+ * Perform work within a launched parallel process
+ */
+void
+HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
+{
+	char	   *sharedquery;
+	HnswSpool  *hnswspool;
+	HnswShared *hnswshared;
+	Relation	heapRel;
+	Relation	indexRel;
+	LOCKMODE	heapLockmode;
+	LOCKMODE	indexLockmode;
+
+	/* Set debug_query_string for individual workers first */
+	sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
+	debug_query_string = sharedquery;
+
+	/* Report the query string from leader */
+	pgstat_report_activity(STATE_RUNNING, debug_query_string);
+
+	/* Look up shared state */
+	hnswshared = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_SHARED, false);
+
+	/* Open relations using lock modes known to be obtained by index.c */
+	if (!hnswshared->isconcurrent)
+	{
+		heapLockmode = ShareLock;
+		indexLockmode = AccessExclusiveLock;
+	}
+	else
+	{
+		heapLockmode = ShareUpdateExclusiveLock;
+		indexLockmode = RowExclusiveLock;
+	}
+
+	/* Open relations within worker */
+#if PG_VERSION_NUM >= 120000
+	heapRel = table_open(hnswshared->heaprelid, heapLockmode);
+#else
+	heapRel = heap_open(hnswshared->heaprelid, heapLockmode);
+#endif
+	indexRel = index_open(hnswshared->indexrelid, indexLockmode);
+
+	/* Initialize worker's own spool */
+	hnswspool = (HnswSpool *) palloc0(sizeof(HnswSpool));
+	hnswspool->heap = heapRel;
+	hnswspool->index = indexRel;
+
+	/* Perform inserts */
+	HnswParallelScanAndInsert(hnswspool, hnswshared, false);
+
+	/* Close relations within worker */
+	index_close(indexRel, indexLockmode);
+#if PG_VERSION_NUM >= 120000
+	table_close(heapRel, heapLockmode);
+#else
+	heap_close(heapRel, heapLockmode);
+#endif
+}
+
+/*
+ * End parallel build
+ */
+static void
+HnswEndParallel(HnswLeader * hnswleader)
+{
+	/* Shutdown worker processes */
+	WaitForParallelWorkersToFinish(hnswleader->pcxt);
+
+	/* Free last reference to MVCC snapshot, if one was used */
+	if (IsMVCCSnapshot(hnswleader->snapshot))
+		UnregisterSnapshot(hnswleader->snapshot);
+	DestroyParallelContext(hnswleader->pcxt);
+	ExitParallelMode();
+}
+
+/*
+ * Return size of shared memory required for parallel index build
+ */
+static Size
+ParallelEstimateShared(Relation heap, Snapshot snapshot)
+{
+#if PG_VERSION_NUM >= 120000
+	return add_size(BUFFERALIGN(sizeof(HnswShared)), table_parallelscan_estimate(heap, snapshot));
+#else
+	if (!IsMVCCSnapshot(snapshot))
+	{
+		Assert(snapshot == SnapshotAny);
+		return sizeof(HnswShared);
+	}
+
+	return add_size(offsetof(HnswShared, heapdesc) +
+					offsetof(ParallelHeapScanDescData, phs_snapshot_data),
+					EstimateSnapshotSpace(snapshot));
+#endif
+}
+
+/*
+ * Within leader, participate as a parallel worker
+ */
+static void
+HnswLeaderParticipateAsWorker(HnswBuildState * buildstate)
+{
+	HnswLeader *hnswleader = buildstate->hnswleader;
+	HnswSpool  *leaderworker;
+
+	/* Allocate memory and initialize private spool */
+	leaderworker = (HnswSpool *) palloc0(sizeof(HnswSpool));
+	leaderworker->heap = buildstate->heap;
+	leaderworker->index = buildstate->index;
+
+	/* Perform work common to all participants */
+	HnswParallelScanAndInsert(leaderworker, hnswleader->hnswshared, true);
+}
+
+/*
+ * Begin parallel build
+ */
+static void
+HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
+{
+	ParallelContext *pcxt;
+	int			scantuplesortstates;
+	Snapshot	snapshot;
+	Size		esthnswshared;
+	HnswShared *hnswshared;
+	HnswLeader *hnswleader = (HnswLeader *) palloc0(sizeof(HnswLeader));
+	bool		leaderparticipates = true;
+	int			querylen;
+
+#ifdef DISABLE_LEADER_PARTICIPATION
+	leaderparticipates = false;
+#endif
+
+	/* Enter parallel mode and create context */
+	EnterParallelMode();
+	Assert(request > 0);
+#if PG_VERSION_NUM >= 120000
+	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request);
+#else
+	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request, true);
+#endif
+
+	scantuplesortstates = leaderparticipates ? request + 1 : request;
+
+	/* Get snapshot for table scan */
+	if (!isconcurrent)
+		snapshot = SnapshotAny;
+	else
+		snapshot = RegisterSnapshot(GetTransactionSnapshot());
+
+	/* Estimate size of workspaces */
+	esthnswshared = ParallelEstimateShared(buildstate->heap, snapshot);
+	shm_toc_estimate_chunk(&pcxt->estimator, esthnswshared);
+	shm_toc_estimate_keys(&pcxt->estimator, 1);
+
+	/* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
+	if (debug_query_string)
+	{
+		querylen = strlen(debug_query_string);
+		shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
+		shm_toc_estimate_keys(&pcxt->estimator, 1);
+	}
+	else
+		querylen = 0;			/* keep compiler quiet */
+
+	/* Everyone's had a chance to ask for space, so now create the DSM */
+	InitializeParallelDSM(pcxt);
+
+	/* If no DSM segment was available, back out (do serial build) */
+	if (pcxt->seg == NULL)
+	{
+		if (IsMVCCSnapshot(snapshot))
+			UnregisterSnapshot(snapshot);
+		DestroyParallelContext(pcxt);
+		ExitParallelMode();
+		return;
+	}
+
+	/* Store shared build state, for which we reserved space */
+	hnswshared = (HnswShared *) shm_toc_allocate(pcxt->toc, esthnswshared);
+	/* Initialize immutable state */
+	hnswshared->heaprelid = RelationGetRelid(buildstate->heap);
+	hnswshared->indexrelid = RelationGetRelid(buildstate->index);
+	hnswshared->isconcurrent = isconcurrent;
+	hnswshared->scantuplesortstates = scantuplesortstates;
+	ConditionVariableInit(&hnswshared->workersdonecv);
+	SpinLockInit(&hnswshared->mutex);
+	/* Initialize mutable state */
+	hnswshared->nparticipantsdone = 0;
+	hnswshared->reltuples = 0;
+	hnswshared->indtuples = 0;
+#if PG_VERSION_NUM >= 120000
+	table_parallelscan_initialize(buildstate->heap,
+								  ParallelTableScanFromHnswShared(hnswshared),
+								  snapshot);
+#else
+	heap_parallelscan_initialize(&hnswshared->heapdesc, buildstate->heap, snapshot);
+#endif
+
+	shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
+
+	/* Store query string for workers */
+	if (debug_query_string)
+	{
+		char	   *sharedquery;
+
+		sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
+		memcpy(sharedquery, debug_query_string, querylen + 1);
+		shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
+	}
+
+	/* Launch workers, saving status for leader/caller */
+	LaunchParallelWorkers(pcxt);
+	hnswleader->pcxt = pcxt;
+	hnswleader->nparticipanttuplesorts = pcxt->nworkers_launched;
+	if (leaderparticipates)
+		hnswleader->nparticipanttuplesorts++;
+	hnswleader->hnswshared = hnswshared;
+	hnswleader->snapshot = snapshot;
+
+	/* If no workers were successfully launched, back out (do serial build) */
+	if (pcxt->nworkers_launched == 0)
+	{
+		HnswEndParallel(hnswleader);
+		return;
+	}
+
+	/* Log participants */
+	ereport(DEBUG1, (errmsg("using %d parallel workers", pcxt->nworkers_launched)));
+
+	/* Save leader state now that it's clear build will be parallel */
+	buildstate->hnswleader = hnswleader;
+
+	/* Join heap scan ourselves */
+	if (leaderparticipates)
+		HnswLeaderParticipateAsWorker(buildstate);
+
+	/* Wait for all launched workers */
+	WaitForParallelWorkersToAttach(pcxt);
+}
+
 /*
  * Build graph
  */
 static void
 BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
 {
+	int			parallel_workers = 0;
+
 	UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_HNSW_PHASE_LOAD);
 
+	/* Calculate parallel workers */
+	if (hnsw_enable_parallel_build)
+		parallel_workers = plan_create_index_workers(RelationGetRelid(buildstate->heap), RelationGetRelid(buildstate->index));
+
+	/* Attempt to launch parallel worker scan when required */
+	if (parallel_workers > 0)
+	{
+		/* TODO Support in-memory builds */
+		FlushPages(buildstate);
+		HnswBeginParallel(buildstate, buildstate->indexInfo->ii_Concurrent, parallel_workers);
+	}
+
+	/* Add tuples to sort */
+	if (buildstate->hnswleader)
+		buildstate->reltuples = ParallelHeapScan(buildstate);
+	else
+	{
 #if PG_VERSION_NUM >= 120000
-	buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+		buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, true, BuildCallback, (void *) buildstate, NULL);
+													   true, true, BuildCallback, (void *) buildstate, NULL);
 #else
-	buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-											   true, BuildCallback, (void *) buildstate, NULL);
+												   true, BuildCallback, (void *) buildstate, NULL);
 #endif
+	}
+
+	/* End parallel build */
+	if (buildstate->hnswleader)
+		HnswEndParallel(buildstate->hnswleader);
 }
 
 /*

src/hnswinsert.c

@@ -123,7 +123,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	Size		minCombinedSize;
 	HnswElementTuple etup;
 	BlockNumber currentPage = insertPage;
-	int			dimensions = e->vec->dim;
 	HnswNeighborTuple ntup;
 	Buffer		nbuf;
 	Page		npage;
@@ -132,7 +131,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -405,8 +404,9 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
+	ItemId		itemid;
 	HnswElementTuple etup;
+	Size		etupSize;
 	int			i;
 
 	/* Read page */
@@ -416,7 +416,9 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	page = GenericXLogRegisterBuffer(state, buf, 0);
 
 	/* Find space */
-	etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
+	itemid = PageGetItemId(page, dup->offno);
+	etup = (HnswElementTuple) PageGetItem(page, itemid);
+	etupSize = ItemIdGetLength(itemid);
 	for (i = 0; i < HNSW_HEAPTIDS; i++)
 	{
 		if (!ItemPointerIsValid(&etup->heaptids[i]))
@@ -498,10 +500,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	if (normprocinfo != NULL)
-	{
+		HnswNormValue(normprocinfo, collation, &value, NULL);
-		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
-			return false;
-	}
 
 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -515,7 +514,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 
 	/* Create an element */
 	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->vec = DatumGetVector(value);
+	element->value = value;
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)

src/hnswutils.c

@@ -4,6 +4,7 @@
 
 #include "hnsw.h"
 #include "storage/bufmgr.h"
+#include "utils/datum.h"
 #include "vector.h"
 
 /*
@@ -54,7 +55,7 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
  * The caller needs to free the pointer stored in value
  * if it's different than the original value
  */
-bool
+void
 HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
@@ -70,11 +71,7 @@ HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
 			result->x[i] = v->x[i] / norm;
 
 		*value = PointerGetDatum(result);
-
-		return true;
 	}
-
-	return false;
 }
 
 /*
@@ -139,6 +136,7 @@ HnswInitNeighbors(HnswElement element, int m)
 		a = &element->neighbors[lc];
 		a->length = 0;
 		a->items = palloc(sizeof(HnswCandidate) * lm);
+		a->closerSet = false;
 	}
 }
 
@@ -175,6 +173,8 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
 
 	HnswInitNeighbors(element, m);
 
+	element->value = PointerGetDatum(NULL);
+
 	return element;
 }
 
@@ -186,7 +186,8 @@ HnswFreeElement(HnswElement element)
 {
 	HnswFreeNeighbors(element);
 	list_free_deep(element->heaptids);
-	pfree(element->vec);
+	if (DatumGetPointer(element->value))
+		pfree(DatumGetPointer(element->value));
 	pfree(element);
 }
 
@@ -213,7 +214,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->vec = NULL;
+	element->value = PointerGetDatum(NULL);
 	return element;
 }
 
@@ -323,7 +324,7 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
+	memcpy(&etup->data, DatumGetPointer(element->value), VARSIZE_ANY(DatumGetPointer(element->value)));
 }
 
 /*
@@ -445,10 +446,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	}
 
 	if (loadVec)
-	{
+		element->value = datumCopy(PointerGetDatum(&etup->data), false, -1);
-		element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
-		memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
-	}
 }
 
 /*
@@ -475,7 +473,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
+		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
 
 	UnlockReleaseBuffer(buf);
 }
@@ -486,7 +484,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 static float
 GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
 {
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
 }
 
 /*
@@ -692,6 +690,34 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 	return w;
 }
 
+/*
+ * Compare candidate distances
+ */
+static int
+#if PG_VERSION_NUM >= 130000
+CompareCandidateDistances(const ListCell *a, const ListCell *b)
+#else
+CompareCandidateDistances(const void *a, const void *b)
+#endif
+{
+	HnswCandidate *hca = lfirst((ListCell *) a);
+	HnswCandidate *hcb = lfirst((ListCell *) b);
+
+	if (hca->distance < hcb->distance)
+		return 1;
+
+	if (hca->distance > hcb->distance)
+		return -1;
+
+	if (hca->element < hcb->element)
+		return 1;
+
+	if (hca->element > hcb->element)
+		return -1;
+
+	return 0;
+}
+
 /*
  * Calculate the distance between elements
  */
@@ -721,7 +747,7 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
 		}
 	}
 
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
 }
 
 /*
@@ -748,33 +774,77 @@ CheckElementCloser(HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid c
  * Algorithm 4 from paper
  */
 static List *
-SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswCandidate * *pruned)
+SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
 	pairingheap *wd;
+	bool		mustCalculate = !e2->neighbors[lc].closerSet;
+	List	   *added = NIL;
+	bool		removedAny = false;
 
 	if (list_length(w) <= m)
 		return w;
 
 	wd = pairingheap_allocate(CompareNearestCandidates, NULL);
 
+	/* Ensure order of candidates is deterministic for closer caching */
+	if (sortCandidates)
+		list_sort(w, CompareCandidateDistances);
+
 	while (list_length(w) > 0 && list_length(r) < m)
 	{
 		/* Assumes w is already ordered desc */
 		HnswCandidate *e = llast(w);
-		bool		closer;
 
 		w = list_delete_last(w);
 
-		closer = CheckElementCloser(e, r, lc, procinfo, collation);
+		/* Use previous state of r and wd to skip work when possible */
+		if (mustCalculate)
+			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+		else if (list_length(added) > 0)
+		{
+			/*
+			 * If the current candidate was closer, we only need to compare it
+			 * with the other candidates that we have added.
+			 */
+			if (e->closer)
+			{
+				e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
 
-		if (closer)
+				if (!e->closer)
+					removedAny = true;
+			}
+			else
+			{
+				/*
+				 * If we have removed any candidates from closer, a candidate
+				 * that was not closer earlier might now be.
+				 */
+				if (removedAny)
+				{
+					e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+					if (e->closer)
+						added = lappend(added, e);
+				}
+			}
+		}
+		else if (e == newCandidate)
+		{
+			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+			if (e->closer)
+				added = lappend(added, e);
+		}
+
+		if (e->closer)
 			r = lappend(r, e);
 		else
 			pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
 	}
 
+	/* Cached value can only be used in future if sorted deterministically */
+	e2->neighbors[lc].closerSet = sortCandidates;
+
 	/* Keep pruned connections */
 	while (!pairingheap_is_empty(wd) && list_length(r) < m)
 		r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -804,7 +874,7 @@ HnswFindDuplicate(HnswElement e)
 		HnswCandidate *neighbor = &neighbors->items[i];
 
 		/* Exit early since ordered by distance */
-		if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
+		if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
 			break;
 
 		/* Check for space */
@@ -828,28 +898,6 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
 		a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
 }
 
-/*
- * Compare candidate distances
- */
-static int
-#if PG_VERSION_NUM >= 130000
-CompareCandidateDistances(const ListCell *a, const ListCell *b)
-#else
-CompareCandidateDistances(const void *a, const void *b)
-#endif
-{
-	HnswCandidate *hca = lfirst((ListCell *) a);
-	HnswCandidate *hcb = lfirst((ListCell *) b);
-
-	if (hca->distance < hcb->distance)
-		return 1;
-
-	if (hca->distance > hcb->distance)
-		return -1;
-
-	return 0;
-}
-
 /*
  * Update connections
  */
@@ -879,13 +927,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = PointerGetDatum(hc->element->vec);
+			Datum		q = hc->element->value;
 
 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (hc3->element->vec == NULL)
+				if (DatumGetPointer(hc3->element->value) == NULL)
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -903,13 +951,12 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		{
 			List	   *c = NIL;
 
-			/* Add and sort candidates */
+			/* Add candidates */
 			for (int i = 0; i < currentNeighbors->length; i++)
 				c = lappend(c, &currentNeighbors->items[i]);
 			c = lappend(c, &hc2);
-			list_sort(c, CompareCandidateDistances);
 
-			SelectNeighbors(c, m, lc, procinfo, collation, &pruned);
+			SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
 
 			/* Should not happen */
 			if (pruned == NULL)
@@ -967,7 +1014,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = PointerGetDatum(element->vec);
+	Datum		q = element->value;
 	HnswElement skipElement = existing ? element : NULL;
 
 	/* No neighbors if no entry point */
@@ -1008,7 +1055,12 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		else
 			lw = w;
 
-		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
+		/*
+		 * Candidates are sorted, but not deterministically. Could set
+		 * sortCandidates to true for in-memory builds to enable closer
+		 * caching, but there does not seem to be a difference in performance.
+		 */
+		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL, false);
 
 		AddConnections(element, neighbors, lm, lc);
 

src/hnswvacuum.c

@@ -62,7 +62,8 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 		/* Iterate over nodes */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			int			idx = 0;
 			bool		itemUpdated = false;
 
@@ -93,7 +94,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 				if (itemUpdated)
 				{
-					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+					Size		etupSize = ItemIdGetLength(itemid);
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -477,7 +478,8 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 		/* Update element and neighbors together */
 		for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
 		{
-			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
+			ItemId		itemid = PageGetItemId(page, offno);
+			HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, itemid);
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
@@ -505,7 +507,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 				continue;
 
 			/* Calculate sizes */
-			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+			etupSize = ItemIdGetLength(itemid);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -528,7 +530,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
+			MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)

src/ivfbuild.c

@@ -177,10 +177,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
-	{
+		IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec);
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
-			return;
-	}
 
 	/* Find the list that minimizes the distance */
 	for (int i = 0; i < centers->length; i++)

src/ivfflat.h

@@ -246,8 +246,6 @@ typedef struct IvfflatScanOpaqueData
 	int			probes;
 	int			dimensions;
 	bool		first;
-	Buffer		buf;
-	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;

src/ivfinsert.c

@@ -83,10 +83,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	/* Normalize if needed */
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
-	{
+		IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL);
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
-			return;
-	}
 
 	/* Find the insert page - sets the page and list info */
 	FindInsertPage(index, values, &insertPage, &listInfo);

src/ivfscan.c

@@ -143,10 +143,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				bool		isnull;
 				ItemId		itemid = PageGetItemId(page, offno);
 
-				/* Skip dead tuples */
-				if (scan->ignore_killed_tuples && ItemIdIsDead(itemid))
-					continue;
-
 				itup = (IndexTuple) PageGetItem(page, itemid);
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
 
@@ -161,8 +157,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
-				slot->tts_values[2] = Int32GetDatum((int) searchPage);
-				slot->tts_isnull[2] = false;
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
@@ -187,55 +181,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
-/*
- * Mark prior tuple as dead
- */
-static void
-MarkPriorTupleDead(IndexScanDesc scan)
-{
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	Buffer		buf = so->buf;
-	Page		page;
-	OffsetNumber maxoffno;
-
-	/* Safety check */
-	if (!BufferIsValid(so->buf) || !ItemPointerIsValid(&so->heaptid))
-		return;
-
-	/* Only a shared locked is needed for ItemIdMarkDead */
-	LockBuffer(buf, BUFFER_LOCK_SHARE);
-	page = BufferGetPage(buf);
-	maxoffno = PageGetMaxOffsetNumber(page);
-
-	for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
-	{
-		ItemId		itemid = PageGetItemId(page, offno);
-		IndexTuple	itup = (IndexTuple) PageGetItem(page, itemid);
-
-		/*
-		 * Find tuple. Since buffer has been pinned, tuple cannot have been
-		 * vacuumed (and heap TID reused).
-		 */
-		if (ItemPointerEquals(&itup->t_tid, &so->heaptid))
-		{
-			/*
-			 * Make sure tuple has not already been marked dead to avoid extra
-			 * WAL if wal_log_hints or data checksums enabled
-			 */
-			if (!ItemIdIsDead(itemid))
-			{
-				ItemIdMarkDead(itemid);
-				MarkBufferDirtyHint(buf, true);
-			}
-
-			break;
-		}
-	}
-
-	/* Unlock buffer */
-	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
-}
-
 /*
  * Prepare for an index scan
  */
@@ -261,9 +206,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 		probes = lists;
 
 	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
-	so->buf = InvalidBuffer;
 	so->first = true;
-	ItemPointerSetInvalid(&so->heaptid);
 	so->probes = probes;
 	so->dimensions = dimensions;
 
@@ -274,13 +217,12 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(3);
+	so->tupdesc = CreateTemplateTupleDesc(2);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(3, false);
+	so->tupdesc = CreateTemplateTupleDesc(2, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -312,7 +254,6 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 #endif
 
 	so->first = true;
-	ItemPointerSetInvalid(&so->heaptid);
 	pairingheap_reset(so->listQueue);
 
 	if (keys && scan->numberOfKeys > 0)
@@ -347,6 +288,11 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
+		/* Requires MVCC-compliant snapshot as not able to pin during sorting */
+		/* https://www.postgresql.org/docs/current/index-locking.html */
+		if (!IsMVCCSnapshot(scan->xs_snapshot))
+			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
+
 		if (scan->orderByData->sk_flags & SK_ISNULL)
 			value = PointerGetDatum(InitVector(so->dimensions));
 		else
@@ -370,17 +316,10 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (value != scan->orderByData->sk_argument)
 			pfree(DatumGetPointer(value));
 	}
-	else
-	{
-		/* Mark prior tuple as dead */
-		if (scan->kill_prior_tuple)
-			MarkPriorTupleDead(scan);
-	}
 
 	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
 	{
 		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
-		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
 
 #if PG_VERSION_NUM >= 120000
 		scan->xs_heaptid = *heaptid;
@@ -388,21 +327,6 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		scan->xs_ctup.t_self = *heaptid;
 #endif
 
-		/* Keep track of info needed to mark tuple as dead */
-		so->heaptid = *heaptid;
-
-		/* Unpin buffer */
-		if (BufferIsValid(so->buf))
-			ReleaseBuffer(so->buf);
-
-		/*
-		 * An index scan must maintain a pin on the index page holding the
-		 * item last returned by amgettuple
-		 *
-		 * https://www.postgresql.org/docs/current/index-locking.html
-		 */
-		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
-
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -418,10 +342,6 @@ ivfflatendscan(IndexScanDesc scan)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 
-	/* Release pin */
-	if (BufferIsValid(so->buf))
-		ReleaseBuffer(so->buf);
-
 	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 

src/tinyint.c

@@ -1,294 +0,0 @@
-#include "postgres.h"
-
-#include <math.h>
-#include <stdint.h>
-
-#include "fmgr.h"
-#include "lib/stringinfo.h"
-#include "libpq/pqformat.h"
-#include "tinyint.h"
-#include "utils/array.h"
-#include "utils/builtins.h"
-#include "utils/numeric.h"
-
-/*
- * Check if array is a vector
- */
-static bool
-ArrayIsVector(ArrayType *a)
-{
-	return ARR_NDIM(a) == 1 && !array_contains_nulls(a);
-}
-
-/*
- * Check if dimensions are the same
- */
-static int
-CheckDims(ArrayType *a, ArrayType *b)
-{
-	int			dima;
-	int			dimb;
-
-	if (!ArrayIsVector(a) || !ArrayIsVector(b))
-		return 0;
-
-	dima = ARR_DIMS(a)[0];
-	dimb = ARR_DIMS(b)[0];
-
-	if (dima != dimb)
-		return 0;
-
-	return dima;
-}
-
-/*
- * Check range
- */
-static void
-CheckRange(long i)
-{
-	if (i < INT8_MIN || i > INT8_MAX)
-		ereport(ERROR,
-				(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-				 errmsg("value \"%ld\" is out of range for type tinyint", i)));
-}
-
-/*
- * Convert textual representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_in);
-Datum
-tinyint_in(PG_FUNCTION_ARGS)
-{
-	char	   *s = PG_GETARG_CSTRING(0);
-	const char *ptr = s;
-	long		i;
-	char	   *end;
-
-	/* skip leading spaces */
-	while (*ptr != '\0' && isspace((unsigned char) *ptr))
-		ptr++;
-
-	if (*ptr == '\0')
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type tinyint: \"%s\"", s)));
-
-	i = strtol(ptr, &end, 10);
-	ptr = end;
-
-	if (i < INT8_MIN || i > INT8_MAX)
-		ereport(ERROR,
-				(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-				 errmsg("value \"%s\" is out of range for type tinyint", s)));
-
-	/* allow trailing whitespace, but not other trailing chars */
-	while (*ptr != '\0' && isspace((unsigned char) *ptr))
-		ptr++;
-
-	if (*ptr != '\0')
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type tinyint: \"%s\"", s)));
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Convert internal representation to textual representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_out);
-Datum
-tinyint_out(PG_FUNCTION_ARGS)
-{
-	int8		num = PG_GETARG_INT8(0);
-	char	   *result = (char *) palloc(5);	/* sign, 3 digits, '\0' */
-
-	pg_ltoa((int32) num, result);
-	PG_RETURN_CSTRING(result);
-}
-
-/*
- * Convert external binary representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_recv);
-Datum
-tinyint_recv(PG_FUNCTION_ARGS)
-{
-	StringInfo	buf = (StringInfo) PG_GETARG_POINTER(0);
-
-	PG_RETURN_INT8((int8) pq_getmsgint(buf, sizeof(int8)));
-}
-
-/*
- * Convert internal representation to the external binary representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_send);
-Datum
-tinyint_send(PG_FUNCTION_ARGS)
-{
-	int8		arg1 = PG_GETARG_INT8(0);
-	StringInfoData buf;
-
-	pq_begintypsend(&buf);
-	pq_sendint8(&buf, arg1);
-	PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
-}
-
-/*
- * Convert integer to tinyint
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(integer_to_tinyint);
-Datum
-integer_to_tinyint(PG_FUNCTION_ARGS)
-{
-	int32		i = PG_GETARG_INT32(0);
-
-	CheckRange(i);
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Convert numeric to tinyint
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(numeric_to_tinyint);
-Datum
-numeric_to_tinyint(PG_FUNCTION_ARGS)
-{
-	Numeric		num = PG_GETARG_NUMERIC(0);
-	int32		i = numeric_int4_opt_error(num, NULL);
-
-	CheckRange(i);
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Get the L2 distance between tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_l2_distance);
-Datum
-tinyint_l2_distance(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8    *ax = (int8 *) ARR_DATA_PTR(a);
-	int8    *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		double		diff = ax[i] - bx[i];
-
-		distance += diff * diff;
-	}
-
-	PG_RETURN_FLOAT8(sqrt(distance));
-}
-
-/*
- * Get the inner product of two tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_inner_product);
-Datum
-tinyint_inner_product(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8	   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8	   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += ax[i] * bx[i];
-
-	PG_RETURN_FLOAT8(distance);
-}
-
-/*
- * Get the negative inner product of two tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_negative_inner_product);
-Datum
-tinyint_negative_inner_product(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8	   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8	   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += ax[i] * bx[i];
-
-	PG_RETURN_FLOAT8(distance * -1);
-}
-
-/*
- * Get the cosine distance between two float2 arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_cosine_distance);
-Datum
-tinyint_cosine_distance(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	double		norma = 0.0;
-	double		normb = 0.0;
-	double		similarity;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		float		axi = ax[i];
-		float		bxi = bx[i];
-
-		distance += axi * bxi;
-		norma += axi * axi;
-		normb += bxi * bxi;
-	}
-
-	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
-	similarity = distance / sqrt(norma * normb);
-
-#ifdef _MSC_VER
-	/* /fp:fast may not propagate NaN */
-	if (isnan(similarity))
-		PG_RETURN_FLOAT8(NAN);
-#endif
-
-	/* Keep in range */
-	if (similarity > 1)
-		similarity = 1;
-	else if (similarity < -1)
-		similarity = -1;
-
-	PG_RETURN_FLOAT8(1 - similarity);
-}

src/tinyint.h

@@ -1,8 +0,0 @@
-#ifndef TINYINT_H
-#define TINYINT_H
-
-#define DatumGetInt8(X) ((int8) (X))
-#define PG_GETARG_INT8(n)	 DatumGetInt8(PG_GETARG_DATUM(n))
-#define PG_RETURN_INT8(x)	 return Int8GetDatum(x)
-
-#endif

src/vector.c

@@ -89,7 +89,7 @@ CheckDim(int dim)
 }
 
 /*
- * Ensure finite elements
+ * Ensure finite element
  */
 static inline void
 CheckElement(float value)
@@ -437,17 +437,18 @@ vector_send(PG_FUNCTION_ARGS)
 
 /*
  * Convert vector to vector
+ * This is needed to check the type modifier
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(vector);
 Datum
 vector(PG_FUNCTION_ARGS)
 {
-	Vector	   *arg = PG_GETARG_VECTOR_P(0);
+	Vector	   *vec = PG_GETARG_VECTOR_P(0);
 	int32		typmod = PG_GETARG_INT32(1);
 
-	CheckExpectedDim(typmod, arg->dim);
+	CheckExpectedDim(typmod, vec->dim);
 
-	PG_RETURN_POINTER(arg);
+	PG_RETURN_POINTER(vec);
 }
 
 /*
@@ -464,7 +465,6 @@ array_to_vector(PG_FUNCTION_ARGS)
 	bool		typbyval;
 	char		typalign;
 	Datum	   *elemsp;
-	bool	   *nullsp;
 	int			nelemsp;
 
 	if (ARR_NDIM(array) > 1)
@@ -478,7 +478,7 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("array must not contain nulls")));
 
 	get_typlenbyvalalign(ARR_ELEMTYPE(array), &typlen, &typbyval, &typalign);
-	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, &nullsp, &nelemsp);
+	deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, NULL, &nelemsp);
 
 	CheckDim(nelemsp);
 	CheckExpectedDim(typmod, nelemsp);
@@ -512,6 +512,12 @@ array_to_vector(PG_FUNCTION_ARGS)
 				 errmsg("unsupported array type")));
 	}
 
+	/*
+	 * Free allocation from deconstruct_array. Do not free individual elements
+	 * when pass-by-reference since they point to original array.
+	 */
+	pfree(elemsp);
+
 	/* Check elements */
 	for (int i = 0; i < result->dim; i++)
 		CheckElement(result->x[i]);

test/expected/functions.out

@@ -54,85 +54,85 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]', '[3,4]');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[0,0]', '[0,1]');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[3e38]', '[-3e38]');
  l2_distance 
 -------------
     Infinity
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
+SELECT inner_product('[1,2]', '[3,4]');
  inner_product 
 ---------------
             11
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3]');
+SELECT inner_product('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[3e38]', '[3e38]');
  inner_product 
 ---------------
       Infinity
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
+SELECT cosine_distance('[1,2]', '[2,4]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
+SELECT cosine_distance('[1,2]', '[0,0]');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
+SELECT cosine_distance('[1,1]', '[1,1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
+SELECT cosine_distance('[1,0]', '[0,2]');
  cosine_distance 
 -----------------
                1
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[3]');
+SELECT cosine_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[3e38]', '[3e38]');
  cosine_distance 
 -----------------
              NaN

test/expected/hnsw_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/ivfflat_cosine.out

@@ -9,18 +9,19 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
  [1,1,1]
  [1,2,3]
  [1,2,4]
-(3 rows)
+ [0,0,0]
+(4 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
  count 
 -------
-     3
+     4
 (1 row)
 
 DROP TABLE t;

test/expected/tinyint.out

@@ -1,148 +0,0 @@
-SELECT '127'::tinyint;
- tinyint 
----------
- 127
-(1 row)
-
-SELECT '128'::tinyint;
-ERROR:  value "128" is out of range for type tinyint
-LINE 1: SELECT '128'::tinyint;
-               ^
-SELECT '-128'::tinyint;
- tinyint 
----------
- -128
-(1 row)
-
-SELECT '-129'::tinyint;
-ERROR:  value "-129" is out of range for type tinyint
-LINE 1: SELECT '-129'::tinyint;
-               ^
-SELECT ''::tinyint;
-ERROR:  invalid input syntax for type tinyint: ""
-LINE 1: SELECT ''::tinyint;
-               ^
-SELECT ' 1'::tinyint;
- tinyint 
----------
- 1
-(1 row)
-
-SELECT '1 '::tinyint;
- tinyint 
----------
- 1
-(1 row)
-
-SELECT '1a'::tinyint;
-ERROR:  invalid input syntax for type tinyint: "1a"
-LINE 1: SELECT '1a'::tinyint;
-               ^
-SELECT '{1,2,3}'::tinyint[];
- tinyint 
----------
- {1,2,3}
-(1 row)
-
-SELECT '128'::numeric::tinyint;
-ERROR:  value "128" is out of range for type tinyint
-SELECT 'NaN'::numeric::tinyint;
-ERROR:  cannot convert NaN to integer
-SELECT l2_distance('{0,0}'::tinyint[], '{3,4}'::tinyint[]);
- l2_distance 
--------------
-           5
-(1 row)
-
-SELECT l2_distance('{0,0}'::tinyint[], '{0,1}'::tinyint[]);
- l2_distance 
--------------
-           1
-(1 row)
-
-SELECT l2_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
- l2_distance 
--------------
-            
-(1 row)
-
-SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]);
-ERROR:  invalid input syntax for type tinyint: "3e38"
-LINE 1: SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]...
-                           ^
-SELECT '{0,0}'::tinyint[] <-> '{3,4}'::tinyint[];
- ?column? 
-----------
-        5
-(1 row)
-
-SELECT inner_product('{1,2}'::tinyint[], '{3,4}'::tinyint[]);
- inner_product 
----------------
-            11
-(1 row)
-
-SELECT inner_product('{1,2}'::tinyint[], '{3}'::tinyint[]);
- inner_product 
----------------
-              
-(1 row)
-
-SELECT inner_product('{127}'::tinyint[], '{127}'::tinyint[]);
- inner_product 
----------------
-         16129
-(1 row)
-
-SELECT '{1,2}'::tinyint[] <#> '{3,4}'::tinyint[];
- ?column? 
-----------
-      -11
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{2,4}'::tinyint[]);
- cosine_distance 
------------------
-               0
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{0,0}'::tinyint[]);
- cosine_distance 
------------------
-             NaN
-(1 row)
-
-SELECT cosine_distance('{1,1}'::tinyint[], '{1,1}'::tinyint[]);
- cosine_distance 
------------------
-               0
-(1 row)
-
-SELECT cosine_distance('{1,0}'::tinyint[], '{0,2}'::tinyint[]);
- cosine_distance 
------------------
-               1
-(1 row)
-
-SELECT cosine_distance('{1,1}'::tinyint[], '{-1,-1}'::tinyint[]);
- cosine_distance 
------------------
-               2
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
- cosine_distance 
------------------
-                
-(1 row)
-
-SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyint[]);
-ERROR:  invalid input syntax for type tinyint: "3e38"
-LINE 1: SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyin...
-                               ^
-SELECT '{1,2}'::tinyint[] <=> '{2,4}'::tinyint[];
- ?column? 
-----------
-        0
-(1 row)
-

test/sql/functions.sql

@@ -13,24 +13,24 @@ SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]', '[3,4]');
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[0,0]', '[0,1]');
-SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[1,2]', '[3]');
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[3e38]', '[-3e38]');
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
+SELECT inner_product('[1,2]', '[3,4]');
-SELECT inner_product('[1,2]'::vector, '[3]');
+SELECT inner_product('[1,2]', '[3]');
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[3e38]', '[3e38]');
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
+SELECT cosine_distance('[1,2]', '[2,4]');
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
+SELECT cosine_distance('[1,2]', '[0,0]');
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
+SELECT cosine_distance('[1,1]', '[1,1]');
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
+SELECT cosine_distance('[1,0]', '[0,2]');
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
-SELECT cosine_distance('[1,2]'::vector, '[3]');
+SELECT cosine_distance('[1,2]', '[3]');
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[3e38]', '[3e38]');
 
 SELECT l1_distance('[0,0]', '[3,4]');
 SELECT l1_distance('[0,0]', '[0,1]');

test/sql/tinyint.sql

@@ -1,34 +0,0 @@
-SELECT '127'::tinyint;
-SELECT '128'::tinyint;
-SELECT '-128'::tinyint;
-SELECT '-129'::tinyint;
-
-SELECT ''::tinyint;
-SELECT ' 1'::tinyint;
-SELECT '1 '::tinyint;
-SELECT '1a'::tinyint;
-
-SELECT '{1,2,3}'::tinyint[];
-
-SELECT '128'::numeric::tinyint;
-SELECT 'NaN'::numeric::tinyint;
-
-SELECT l2_distance('{0,0}'::tinyint[], '{3,4}'::tinyint[]);
-SELECT l2_distance('{0,0}'::tinyint[], '{0,1}'::tinyint[]);
-SELECT l2_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]);
-SELECT '{0,0}'::tinyint[] <-> '{3,4}'::tinyint[];
-
-SELECT inner_product('{1,2}'::tinyint[], '{3,4}'::tinyint[]);
-SELECT inner_product('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT inner_product('{127}'::tinyint[], '{127}'::tinyint[]);
-SELECT '{1,2}'::tinyint[] <#> '{3,4}'::tinyint[];
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{2,4}'::tinyint[]);
-SELECT cosine_distance('{1,2}'::tinyint[], '{0,0}'::tinyint[]);
-SELECT cosine_distance('{1,1}'::tinyint[], '{1,1}'::tinyint[]);
-SELECT cosine_distance('{1,0}'::tinyint[], '{0,2}'::tinyint[]);
-SELECT cosine_distance('{1,1}'::tinyint[], '{-1,-1}'::tinyint[]);
-SELECT cosine_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyint[]);
-SELECT '{1,2}'::tinyint[] <=> '{2,4}'::tinyint[];

test/t/012_hnsw_build_recall.pl

@@ -83,11 +83,32 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Add index
+	# Build index serially
-	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
+	$node->safe_psql("postgres", qq(
+		SET max_parallel_maintenance_workers = 0;
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
 
+	# Test approximate results
 	my $min = $operator eq "<#>" ? 0.80 : 0.99;
 	test_recall($min, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+
+	# Build index in parallel
+	my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+		SET client_min_messages = DEBUG;
+		SET min_parallel_table_scan_size = 1;
+		SET hnsw.enable_parallel_build = on;
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
+	is($ret, 0, $stderr);
+	like($stderr, qr/using \d+ parallel workers/);
+
+	# Test approximate results
+	test_recall($min, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
 }
 
 done_testing();

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.5.0'
+default_version = '0.5.1'
 module_pathname = '$libdir/vector'
 relocatable = true