Less space [skip ci]

Use consistent order [skip ci]
Added InitVisited [skip ci]
2026-07-22 20:15:46 +08:00 · 2024-01-22 18:43:26 -08:00 · 2024-01-22 18:39:52 -08:00 · 2024-01-22 18:37:37 -08:00 · 2024-01-22 18:02:24 -08:00 · 2024-01-22 17:58:42 -08:00
13 changed files with 1154 additions and 531 deletions
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -62,7 +62,7 @@ jobs:
          wget -q https://github.com/postgres/postgres/archive/refs/tags/REL_14_5.tar.gz
          tar xf REL_14_5.tar.gz
      - run: make prove_installcheck PROVE_FLAGS="-I ./postgres-REL_14_5/src/test/perl" PERL5LIB="/Users/runner/perl5/lib/perl5"
-      - run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make
+      - run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make PG_CFLAGS="-DUSE_ASSERT_CHECKING"
  windows:
    runs-on: windows-latest
    if: ${{ !startsWith(github.ref_name, 'mac') }}
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,7 +1,8 @@
 ## 0.5.2 (unreleased)

 - Improved performance of HNSW
- Added support for on-disk parallel index builds for HNSW
+- Added support for parallel index builds for HNSW
+- Reduced memory usage for HNSW index builds
 - Reduced WAL generation for HNSW index builds
 - Fixed `invalid memory alloc request size` error with HNSW index build

--- a/README.md
+++ b/README.md
@@ -161,8 +161,97 @@ You can add an index to use approximate nearest neighbor search, which trades so

 Supported index types are:

- [IVFFlat](#ivfflat)
 - [HNSW](#hnsw) - added in 0.5.0
+- [IVFFlat](#ivfflat)
+
+## HNSW
+
+An HNSW index creates a multilayer graph. It has better query performance than IVFFlat (in terms of speed-recall tradeoff), but has slower build times and uses more memory. Also, an index can be created without any data in the table since there isn’t a training step like IVFFlat.
+
+Add an index for each distance function you want to use.
+
+L2 distance
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
+```
+
+Inner product
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
+```
+
+Cosine distance
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
+```
+
+Vectors with up to 2,000 dimensions can be indexed.
+
+### Index Options
+
+Specify HNSW parameters
+
+- `m` - the max number of connections per layer (16 by default)
+- `ef_construction` - the size of the dynamic candidate list for constructing the graph (64 by default)
+
+```sql
+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)
+
+```sql
+SET hnsw.ef_search = 100;
+```
+
+A higher value provides better recall at the cost of speed.
+
+Use `SET LOCAL` inside a transaction to set it for a single query
+
+```sql
+BEGIN;
+SET LOCAL hnsw.ef_search = 100;
+SELECT ...
+COMMIT;
+```
+
+### Index Build Time
+
+Indexes build significantly faster when the graph fits into `maintenance_work_mem`
+
+```sql
+SET maintenance_work_mem = '8GB';
+```
+
+A notice is shown when the graph no longer fits
+
+```text
+NOTICE:  hnsw graph no longer fits into maintenance_work_mem after 100000 tuples
+DETAIL:  Building will take significantly more time.
+HINT:  Increase maintenance_work_mem to speed up builds.
+```
+
+Note: Do not set `maintenance_work_mem` so high that it exhausts the memory on the server
+
+### 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 * blocks_done / nullif(blocks_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
+```
+
+The phases for HNSW are:
+
+1. `initializing`
+2. `loading tuples`

 ## IVFFlat

@@ -232,77 +321,6 @@ The phases for IVFFlat are:

 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.
-
-Add an index for each distance function you want to use.
-
-L2 distance
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);
-```
-
-Inner product
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);
-```
-
-Cosine distance
-
-```sql
-CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
-```
-
-Vectors with up to 2,000 dimensions can be indexed.
-
-### Index Options
-
-Specify HNSW parameters
-
- `m` - the max number of connections per layer (16 by default)
- `ef_construction` - the size of the dynamic candidate list for constructing the graph (64 by default)
-
-```sql
-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)
-
-```sql
-SET hnsw.ef_search = 100;
-```
-
-A higher value provides better recall at the cost of speed.
-
-Use `SET LOCAL` inside a transaction to set it for a single query
-
-```sql
-BEGIN;
-SET LOCAL hnsw.ef_search = 100;
-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 * blocks_done / nullif(blocks_total, 0), 1) AS "%" FROM pg_stat_progress_create_index;
-```
-
-The phases for HNSW are:
-
-1. `initializing`
-2. `loading tuples`
-
 ## Filtering

 There are a few ways to index nearest neighbor queries with a `WHERE` clause
@@ -320,8 +338,7 @@ CREATE INDEX ON items (category_id);
 Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search

 ```sql
-CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)
-    WHERE (category_id = 123);
+CREATE INDEX ON items USING hnsw (embedding vector_l2_ops) WHERE (category_id = 123);
 ```

 Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html) for approximate search on many different values of the `WHERE` columns
@@ -571,10 +588,10 @@ If compilation fails with `fatal error: postgres.h: No such file or directory`,
 For Ubuntu and Debian, use:

 ```sh
-sudo apt install postgresql-server-dev-15
+sudo apt install postgresql-server-dev-16
 ```

-Note: Replace `15` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### Windows

@@ -589,7 +606,7 @@ Note: The exact path will vary depending on your Visual Studio version and editi
 Then use `nmake` to build:

 ```cmd
-set "PGROOT=C:\Program Files\PostgreSQL\15"
+set "PGROOT=C:\Program Files\PostgreSQL\16"
 git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
@@ -639,22 +656,22 @@ pgxn install vector
 Debian and Ubuntu packages are available from the [PostgreSQL APT Repository](https://wiki.postgresql.org/wiki/Apt). Follow the [setup instructions](https://wiki.postgresql.org/wiki/Apt#Quickstart) and run:

 ```sh
-sudo apt install postgresql-15-pgvector
+sudo apt install postgresql-16-pgvector
 ```

-Note: Replace `15` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### Yum

 RPM packages are available from the [PostgreSQL Yum Repository](https://yum.postgresql.org/). Follow the [setup instructions](https://www.postgresql.org/download/linux/redhat/) for your distribution and run:

 ```sh
-sudo yum install pgvector_15
+sudo yum install pgvector_16
 # or
-sudo dnf install pgvector_15
+sudo dnf install pgvector_16
 ```

-Note: Replace `15` with your Postgres server version
+Note: Replace `16` with your Postgres server version

 ### conda-forge

--- a/src/hnsw.c
+++ b/src/hnsw.c
@@ -14,15 +14,45 @@
 #endif

 int			hnsw_ef_search;
-bool		hnsw_enable_parallel_build;
+int			hnsw_lock_tranche_id;
 static relopt_kind hnsw_relopt_kind;

+/*
+ * Assign a tranche ID for our LWLocks. This only needs to be done by one
+ * backend, as the tranche ID is remembered in shared memory.
+ *
+ * This shared memory area is very small, so we just allocate it from the
+ * "slop" that PostgreSQL reserves for small allocations like this. If
+ * this grows bigger, we should use a shmem_request_hook and
+ * RequestAddinShmemSpace() to pre-reserve space for this.
+ */
+static void
+HnswInitLockTranche(void)
+{
+	int		   *tranche_ids;
+	bool		found;
+
+	LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
+	tranche_ids = ShmemInitStruct("hnsw LWLock ids",
+								  sizeof(int) * 1,
+								  &found);
+	if (!found)
+		tranche_ids[0] = LWLockNewTrancheId();
+	hnsw_lock_tranche_id = tranche_ids[0];
+	LWLockRelease(AddinShmemInitLock);
+
+	/* Per-backend registration of the tranche ID */
+	LWLockRegisterTranche(hnsw_lock_tranche_id, "HnswBuild");
+}
+
 /*
 * Initialize index options and variables
 */
 void
 HnswInit(void)
 {
+	HnswInitLockTranche();
+
 	hnsw_relopt_kind = add_reloption_kind();
 	add_int_reloption(hnsw_relopt_kind, "m", "Max number of connections",
 					  HNSW_DEFAULT_M, HNSW_MIN_M, HNSW_MAX_M
@@ -40,11 +70,6 @@ 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);
 }

 /*
--- a/src/hnsw.h
+++ b/src/hnsw.h
@@ -8,6 +8,7 @@
 #include "access/reloptions.h"
 #include "nodes/execnodes.h"
 #include "port.h"				/* for random() */
+#include "utils/relptr.h"
 #include "utils/sampling.h"
 #include "vector.h"

@@ -72,8 +73,10 @@

 #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)
 #else
 #define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
+#define SeedRandom(seed) srandom(seed)
 #endif

 #if PG_VERSION_NUM < 130000
@@ -91,33 +94,64 @@
 #define HnswGetMl(m) (1 / log(m))

 /* Ensure fits on page and in uint8 */
-#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / m) - 2, 255)
+#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / (m)) - 2, 255)
+
+#define HnswGetValue(base, element) PointerGetDatum(HnswPtrAccess(base, (element)->value))
+
+#if PG_VERSION_NUM < 140005
+#define relptr_offset(rp) ((rp).relptr_off - 1)
+#endif
+
+/* Pointer macros */
+#define HnswPtrAccess(base, hp) ((base) == NULL ? (hp).ptr : relptr_access(base, (hp).relptr))
+#define HnswPtrStore(base, hp, value) ((base) == NULL ? (void) ((hp).ptr = (value)) : (void) relptr_store(base, (hp).relptr, value))
+#define HnswPtrIsNull(base, hp) ((base) == NULL ? (hp).ptr == NULL : relptr_is_null((hp).relptr))
+#define HnswPtrEqual(base, hp1, hp2) ((base) == NULL ? (hp1).ptr == (hp2).ptr : relptr_offset((hp1).relptr) == relptr_offset((hp2).relptr))
+
+/* For code paths dedicated to each type */
+#define HnswPtrPointer(hp) (hp).ptr
+#define HnswPtrOffset(hp) relptr_offset((hp).relptr)

 /* Variables */
 extern int	hnsw_ef_search;
-extern bool hnsw_enable_parallel_build;
+extern int	hnsw_lock_tranche_id;

+typedef struct HnswElementData HnswElementData;
 typedef struct HnswNeighborArray HnswNeighborArray;

+#define HnswPtrDeclare(type, relptrtype, ptrtype) \
+	relptr_declare(type, relptrtype); \
+	typedef union { type *ptr; relptrtype relptr; } ptrtype;
+
+/* Pointers that can be absolute or relative */
+/* Use char for DatumPtr so works with Pointer */
+HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
+HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
+HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
+HnswPtrDeclare(char, DatumRelptr, DatumPtr);
+
 typedef struct HnswElementData
 {
-	List	   *heaptids;
+	HnswElementPtr next;
+	ItemPointerData heaptids[HNSW_HEAPTIDS];
+	uint8		heaptidsLength;
 	uint8		level;
 	uint8		deleted;
 	uint32		hash;
-	HnswNeighborArray *neighbors;
+	HnswNeighborsPtr neighbors;
 	BlockNumber blkno;
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Datum		value;
+	DatumPtr	value;
+	LWLock		lock;
 }			HnswElementData;

 typedef HnswElementData * HnswElement;

 typedef struct HnswCandidate
 {
-	HnswElement element;
+	HnswElementPtr element;
 	float		distance;
 	bool		closer;
 }			HnswCandidate;
@@ -126,7 +160,7 @@ typedef struct HnswNeighborArray
 {
 	int			length;
 	bool		closerSet;
-	HnswCandidate *items;
+	HnswCandidate items[FLEXIBLE_ARRAY_MEMBER];
 }			HnswNeighborArray;

 typedef struct HnswPairingHeapNode
@@ -143,11 +177,26 @@ typedef struct HnswOptions
 	int			efConstruction; /* size of dynamic candidate list */
 }			HnswOptions;

-typedef struct HnswSpool
+typedef struct HnswGraph
 {
-	Relation	heap;
-	Relation	index;
-}			HnswSpool;
+	/* Graph state */
+	slock_t		lock;
+	HnswElementPtr head;
+	double		indtuples;
+
+	/* Entry state */
+	LWLock		entryLock;
+	HnswElementPtr entryPoint;
+
+	/* Allocations state */
+	LWLock		allocatorLock;
+	long		memoryUsed;
+	long		memoryTotal;
+
+	/* Flushed state */
+	LWLock		flushLock;
+	bool		flushed;
+}			HnswGraph;

 typedef struct HnswShared
 {
@@ -155,7 +204,6 @@ typedef struct HnswShared
 	Oid			heaprelid;
 	Oid			indexrelid;
 	bool		isconcurrent;
-	int			scantuplesortstates;

 	/* Worker progress */
 	ConditionVariable workersdonecv;
@@ -166,7 +214,7 @@ typedef struct HnswShared
 	/* Mutable state */
 	int			nparticipantsdone;
 	double		reltuples;
-	double		indtuples;
+	HnswGraph	graphData;

 #if PG_VERSION_NUM < 120000
 	ParallelHeapScanDescData heapdesc;	/* must come last */
@@ -184,8 +232,15 @@ typedef struct HnswLeader
 	int			nparticipanttuplesorts;
 	HnswShared *hnswshared;
 	Snapshot	snapshot;
+	char	   *hnswarea;
 }			HnswLeader;

+typedef struct HnswAllocator
+{
+	void	   *(*alloc) (Size size, void *state);
+	void	   *state;
+}			HnswAllocator;
+
 typedef struct HnswBuildState
 {
 	/* Info */
@@ -209,20 +264,21 @@ typedef struct HnswBuildState
 	Oid			collation;

 	/* Variables */
-	List	   *elements;
-	HnswElement entryPoint;
+	HnswGraph	graphData;
+	HnswGraph  *graph;
 	double		ml;
 	int			maxLevel;
-	long		memoryLeft;
-	bool		flushed;
 	Vector	   *normvec;

 	/* Memory */
+	MemoryContext graphCtx;
 	MemoryContext tmpCtx;
+	HnswAllocator allocator;

 	/* Parallel builds */
 	HnswLeader *hnswleader;
 	HnswShared *hnswshared;
+	char	   *hnswarea;
 }			HnswBuildState;

 typedef struct HnswMetaPageData
@@ -321,25 +377,24 @@ bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * re
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);
 void		HnswInit(void);
-List	   *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
+List	   *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
 HnswElement HnswGetEntryPoint(Relation index);
 void		HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
-HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
-void		HnswFreeElement(HnswElement element);
+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		HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
-HnswElement HnswFindDuplicate(HnswElement e);
-HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
+void		HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
+HnswCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
 void		HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
-void		HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m);
+void		HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
 void		HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
-void		HnswInitNeighbors(HnswElement element, int m);
-bool		HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building);
-void		HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building);
+void		HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * alloc);
+bool		HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building);
+void		HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building);
 void		HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
 void		HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec);
-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		HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
+void		HnswUpdateConnection(char *base, 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);

@@ -359,6 +414,16 @@ void		hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys,
 bool		hnswgettuple(IndexScanDesc scan, ScanDirection dir);
 void		hnswendscan(IndexScanDesc scan);

+static inline HnswNeighborArray *
+HnswGetNeighbors(char *base, HnswElement element, int lc)
+{
+	HnswNeighborArrayPtr *neighborList = HnswPtrAccess(base, element->neighbors);
+
+	Assert(element->level >= lc);
+
+	return HnswPtrAccess(base, neighborList[lc]);
+}
+
 /* Hash tables */
 typedef struct TidHashEntry
 {
@@ -386,4 +451,17 @@ typedef struct PointerHashEntry
 #define SH_DECLARE
 #include "lib/simplehash.h"

+typedef struct OffsetHashEntry
+{
+	Size		offset;
+	char		status;
+}			OffsetHashEntry;
+
+#define SH_PREFIX offsethash
+#define SH_ELEMENT_TYPE OffsetHashEntry
+#define SH_KEY_TYPE Size
+#define SH_SCOPE extern
+#define SH_DECLARE
+#include "lib/simplehash.h"
+
 #endif
--- a/src/hnswbuild.c
+++ b/src/hnswbuild.c
@@ -1,3 +1,39 @@
+/*
+ * The HNSW build happens in two phases:
+ *
+ * 1. In-memory phase
+ *
+ * In this first phase, the graph is held completely in memory. When the graph
+ * is fully built, or we run out of memory reserved for the build (determined
+ * by maintenance_work_mem), we materialize the graph to disk (see
+ * FlushPages()), and switch to the on-disk phase.
+ *
+ * In a parallel build, a large contiguous chunk of shared memory is allocated
+ * to hold the graph. Each worker process has its own HnswBuildState struct in
+ * private memory, which contains information that doesn't change throughout
+ * the build, and pointers to the shared structs in shared memory. The shared
+ * memory area is mapped to a different address in each worker process, and
+ * 'HnswBuildState.hnswarea' points to the beginning of the shared area in the
+ * worker process's address space. All pointers used in the graph are
+ * "relative pointers", stored as an offset from 'hnswarea'.
+ *
+ * Each element is protected by an LWLock. It must be held when reading or
+ * modifying the element's neighbors or 'heaptids'.
+ *
+ * In a non-parallel build, the graph is held in backend-private memory. All
+ * the elements are allocated in a dedicated memory context, 'graphCtx', and
+ * the pointers used in the graph are regular pointers.
+ *
+ * 2. On-disk phase
+ *
+ * In the on-disk phase, the index is built by inserting each vector to the
+ * index one by one, just like on INSERT. The only difference is that we don't
+ * WAL-log the individual inserts. If the graph fit completely in memory and
+ * was fully built in the in-memory phase, the on-disk phase is skipped.
+ *
+ * After we have finished building the graph, we perform one more scan through
+ * the index and write all the pages to the WAL.
+ */
 #include "postgres.h"

 #include <math.h>
@@ -54,9 +90,12 @@
 #endif

 #define PARALLEL_KEY_HNSW_SHARED		UINT64CONST(0xA000000000000001)
-#define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000002)
+#define PARALLEL_KEY_HNSW_AREA			UINT64CONST(0xA000000000000002)
+#define PARALLEL_KEY_QUERY_TEXT			UINT64CONST(0xA000000000000003)

-#define LIST_MAX_LENGTH ((1 << 26) - 1)
+#if PG_VERSION_NUM < 130000
+#define GENERATIONCHUNK_RAWSIZE (SIZEOF_SIZE_T + SIZEOF_VOID_P * 2)
+#endif

 /*
 * Create the metapage
@@ -88,6 +127,7 @@ CreateMetaPage(HnswBuildState * buildstate)
 	((PageHeader) page)->pd_lower =
 		((char *) metap + sizeof(HnswMetaPageData)) - (char *) page;

+	MarkBufferDirty(buf);
 	UnlockReleaseBuffer(buf);
 }

@@ -104,6 +144,7 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, ForkNumber forkNum)
 	HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);

 	/* Commit */
+	MarkBufferDirty(*buf);
 	UnlockReleaseBuffer(*buf);

 	/* Can take a while, so ensure we can interrupt */
@@ -131,9 +172,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
 	BlockNumber insertPage;
+	HnswElement entryPoint;
 	Buffer		buf;
 	Page		page;
-	ListCell   *lc;
+	HnswElementPtr iter = buildstate->graph->head;
+	char	   *base = buildstate->hnswarea;

 	/* Calculate sizes */
 	etupAllocSize = BLCKSZ;
@@ -148,18 +191,22 @@ CreateElementPages(HnswBuildState * buildstate)
 	page = BufferGetPage(buf);
 	HnswInitPage(buf, page);

-	foreach(lc, buildstate->elements)
+	while (!HnswPtrIsNull(base, iter))
 	{
-		HnswElement element = lfirst(lc);
+		HnswElement element = HnswPtrAccess(base, iter);
 		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
+		void	   *valuePtr = HnswPtrAccess(base, element->value);
+
+		/* Update iterator */
+		iter = element->next;

 		/* Zero memory for each element */
 		MemSet(etup, 0, etupAllocSize);

 		/* Calculate sizes */
-		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(element->value)));
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(valuePtr));
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);

@@ -167,7 +214,7 @@ CreateElementPages(HnswBuildState * buildstate)
 		if (etupSize > etupAllocSize)
 			elog(ERROR, "index tuple too large");

-		HnswSetElementTuple(etup, element);
+		HnswSetElementTuple(base, etup, element);

 		/* Keep element and neighbors on the same page if possible */
 		if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
@@ -205,9 +252,11 @@ CreateElementPages(HnswBuildState * buildstate)
 	insertPage = BufferGetBlockNumber(buf);

 	/* Commit */
+	MarkBufferDirty(buf);
 	UnlockReleaseBuffer(buf);

-	HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, buildstate->entryPoint, insertPage, forkNum, true);
+	entryPoint = HnswPtrAccess(base, buildstate->graph->entryPoint);
+	HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, entryPoint, insertPage, forkNum, true);

 	pfree(etup);
 	pfree(ntup);
@@ -222,19 +271,23 @@ CreateNeighborPages(HnswBuildState * buildstate)
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
 	int			m = buildstate->m;
-	ListCell   *lc;
+	HnswElementPtr iter = buildstate->graph->head;
+	char	   *base = buildstate->hnswarea;
 	HnswNeighborTuple ntup;

 	/* Allocate once */
 	ntup = palloc0(BLCKSZ);

-	foreach(lc, buildstate->elements)
+	while (!HnswPtrIsNull(base, iter))
 	{
-		HnswElement e = lfirst(lc);
+		HnswElement e = HnswPtrAccess(base, iter);
 		Buffer		buf;
 		Page		page;
 		Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);

+		/* Update iterator */
+		iter = e->next;
+
 		/* Can take a while, so ensure we can interrupt */
 		/* Needs to be called when no buffer locks are held */
 		CHECK_FOR_INTERRUPTS();
@@ -243,58 +296,203 @@ CreateNeighborPages(HnswBuildState * buildstate)
 		LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
 		page = BufferGetPage(buf);

-		HnswSetNeighborTuple(ntup, e, m);
+		HnswSetNeighborTuple(base, ntup, e, m);

 		if (!PageIndexTupleOverwrite(page, e->neighborOffno, (Item) ntup, ntupSize))
 			elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));

 		/* Commit */
+		MarkBufferDirty(buf);
 		UnlockReleaseBuffer(buf);
 	}

 	pfree(ntup);
 }

-/*
- * Free elements
- */
-static void
-FreeElements(HnswBuildState * buildstate)
-{
-	ListCell   *lc;
-
-	foreach(lc, buildstate->elements)
-		HnswFreeElement(lfirst(lc));
-
-	list_free(buildstate->elements);
-}
-
 /*
 * Flush pages
 */
 static void
 FlushPages(HnswBuildState * buildstate)
 {
+#ifdef HNSW_MEMORY
+	elog(INFO, "memory: %zu MB", buildstate->graph->memoryUsed / (1024 * 1024));
+#endif
+
 	CreateMetaPage(buildstate);
 	CreateElementPages(buildstate);
 	CreateNeighborPages(buildstate);

-	buildstate->flushed = true;
-	FreeElements(buildstate);
+	buildstate->graph->flushed = true;
+	MemoryContextReset(buildstate->graphCtx);
+}
+
+/*
+ * Add a heap TID to an existing element
+ */
+static bool
+AddDuplicateInMemory(HnswElement element, HnswElement dup)
+{
+	LWLockAcquire(&dup->lock, LW_EXCLUSIVE);
+
+	if (dup->heaptidsLength == HNSW_HEAPTIDS)
+	{
+		LWLockRelease(&dup->lock);
+		return false;
+	}
+
+	HnswAddHeapTid(dup, &element->heaptids[0]);
+
+	LWLockRelease(&dup->lock);
+
+	return true;
+}
+
+/*
+ * Find duplicate element
+ */
+static bool
+FindDuplicateInMemory(char *base, HnswElement element)
+{
+	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
+	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 (!datumIsEqual(value, neighborValue, false, -1))
+			return false;
+
+		/* Check for space */
+		if (AddDuplicateInMemory(element, neighborElement))
+			return true;
+	}
+
+	return false;
+}
+
+/*
+ * Add to element list
+ */
+static void
+AddElementInMemory(char *base, HnswGraph * graph, HnswElement element)
+{
+	SpinLockAcquire(&graph->lock);
+	element->next = graph->head;
+	HnswPtrStore(base, graph->head, element);
+	SpinLockRelease(&graph->lock);
+}
+
+/*
+ * Update neighbors
+ */
+static void
+UpdateNeighborsInMemory(char *base, FmgrInfo *procinfo, Oid collation, HnswElement e, int m)
+{
+	for (int lc = e->level; lc >= 0; lc--)
+	{
+		int			lm = HnswGetLayerM(m, lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(base, e, lc);
+
+		for (int i = 0; i < neighbors->length; i++)
+		{
+			HnswCandidate *hc = &neighbors->items[i];
+			HnswElement neighborElement = HnswPtrAccess(base, hc->element);
+
+			/* Keep scan-build happy on Mac x86-64 */
+			Assert(neighborElement);
+
+			/* Use element for lock instead of hc since hc can be replaced */
+			LWLockAcquire(&neighborElement->lock, LW_EXCLUSIVE);
+			HnswUpdateConnection(base, e, hc, lm, lc, NULL, NULL, procinfo, collation);
+			LWLockRelease(&neighborElement->lock);
+		}
+	}
+}
+
+/*
+ * Update graph in memory
+ */
+static void
+UpdateGraphInMemory(FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, HnswBuildState * buildstate)
+{
+	HnswGraph  *graph = buildstate->graph;
+	char	   *base = buildstate->hnswarea;
+
+	/* Look for duplicate */
+	if (FindDuplicateInMemory(base, element))
+		return;
+
+	/* Add element */
+	AddElementInMemory(base, graph, element);
+
+	/* Update neighbors */
+	UpdateNeighborsInMemory(base, procinfo, collation, element, m);
+
+	/* Update entry point if needed (already have lock) */
+	if (entryPoint == NULL || element->level > entryPoint->level)
+		HnswPtrStore(base, graph->entryPoint, element);
+}
+
+/*
+ * Insert tuple in memory
+ */
+static void
+InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
+{
+	FmgrInfo   *procinfo = buildstate->procinfo;
+	Oid			collation = buildstate->collation;
+	HnswGraph  *graph = buildstate->graph;
+	HnswElement entryPoint;
+	LWLock	   *entryLock = &graph->entryLock;
+	int			efConstruction = buildstate->efConstruction;
+	int			m = buildstate->m;
+	char	   *base = buildstate->hnswarea;
+
+	/* Get entry point */
+	LWLockAcquire(entryLock, LW_SHARED);
+	entryPoint = HnswPtrAccess(base, graph->entryPoint);
+
+	/* Prevent concurrent inserts when likely updating entry point */
+	if (entryPoint == NULL || element->level > entryPoint->level)
+	{
+		/* Release shared lock */
+		LWLockRelease(entryLock);
+
+		/* Get exclusive lock */
+		LWLockAcquire(entryLock, LW_EXCLUSIVE);
+
+		/* Get latest entry point after lock is acquired */
+		entryPoint = HnswPtrAccess(base, graph->entryPoint);
+	}
+
+	/* Find neighbors for element */
+	HnswFindElementNeighbors(base, element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
+
+	/* Update graph in memory */
+	UpdateGraphInMemory(procinfo, collation, element, m, efConstruction, entryPoint, buildstate);
+
+	/* Release entry lock */
+	LWLockRelease(entryLock);
 }

 /*
 * Insert tuple
 */
 static bool
-InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState * buildstate, HnswElement * dup, MemoryContext outerCtx)
+InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, HnswBuildState * buildstate)
 {
-	FmgrInfo   *procinfo = buildstate->procinfo;
-	Oid			collation = buildstate->collation;
-	HnswElement entryPoint = buildstate->entryPoint;
-	int			efConstruction = buildstate->efConstruction;
-	int			m = buildstate->m;
-	MemoryContext oldCtx;
+	HnswGraph  *graph = buildstate->graph;
+	HnswElement element;
+	HnswAllocator *allocator = &buildstate->allocator;
+	Size		valueSize;
+	Pointer		valuePtr;
+	LWLock	   *flushLock = &graph->flushLock;
+	char	   *base = buildstate->hnswarea;

 	/* Detoast once for all calls */
 	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -302,56 +500,81 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
+		if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
 			return false;
 	}

-	/* Copy value to element so accessible outside of memory context */
-	oldCtx = MemoryContextSwitchTo(outerCtx);
-	element->value = datumCopy(value, false, -1);
-	MemoryContextSwitchTo(oldCtx);
+	/* Get datum size */
+	valueSize = VARSIZE_ANY(DatumGetPointer(value));

-	/* Insert element in graph */
-	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
+	/* Ensure graph not flushed when inserting */
+	LWLockAcquire(flushLock, LW_SHARED);

-	/* Look for duplicate */
-	*dup = HnswFindDuplicate(element);
-
-	/* Update neighbors if needed */
-	if (*dup == NULL)
+	/* Are we in the on-disk phase? */
+	if (graph->flushed)
 	{
-		for (int lc = element->level; lc >= 0; lc--)
-		{
-			int			lm = HnswGetLayerM(m, lc);
-			HnswNeighborArray *neighbors = &element->neighbors[lc];
+		LWLockRelease(flushLock);

-			for (int i = 0; i < neighbors->length; i++)
-				HnswUpdateConnection(element, &neighbors->items[i], lm, lc, NULL, NULL, procinfo, collation);
-		}
+		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, buildstate->heap, true);
 	}

-	/* Update entry point if needed */
-	if (*dup == NULL && (entryPoint == NULL || element->level > entryPoint->level))
-		buildstate->entryPoint = element;
+	/*
+	 * In a parallel build, the HnswElement is allocated from the shared
+	 * memory area, so we need to coordinate with other processes.
+	 */
+	LWLockAcquire(&graph->allocatorLock, LW_EXCLUSIVE);

-	UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++buildstate->indtuples);
+	/*
+	 * 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)
+	{
+		LWLockRelease(&graph->allocatorLock);

-	return *dup == NULL;
-}
+		LWLockRelease(flushLock);
+		LWLockAcquire(flushLock, LW_EXCLUSIVE);

-/*
- * Get the memory used by an element
- */
-static long
-HnswElementMemory(HnswElement e, int m)
-{
-	long		elementSize = sizeof(HnswElementData);
+		if (!graph->flushed)
+		{
+			ereport(NOTICE,
+					(errmsg("hnsw graph no longer fits into maintenance_work_mem after " INT64_FORMAT " tuples", (int64) graph->indtuples),
+					 errdetail("Building will take significantly more time."),
+					 errhint("Increase maintenance_work_mem to speed up builds.")));

-	elementSize += sizeof(HnswNeighborArray) * (e->level + 1);
-	elementSize += sizeof(HnswCandidate) * (m * (e->level + 2));
-	elementSize += sizeof(ItemPointerData);
-	elementSize += VARSIZE_ANY(DatumGetPointer(e->value));
-	return elementSize;
+			FlushPages(buildstate);
+		}
+
+		LWLockRelease(flushLock);
+
+		return HnswInsertTupleOnDisk(index, value, values, isnull, heaptid, buildstate->heap, true);
+	}
+
+	/* Ok, we can proceed to allocate the element */
+	element = HnswInitElement(base, heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel, allocator);
+	valuePtr = HnswAlloc(allocator, valueSize);
+
+	/*
+	 * We have now allocated the space needed for the element, so we don't
+	 * need the allocator lock anymore. Release it and initialize the rest of
+	 * the element.
+	 */
+	LWLockRelease(&graph->allocatorLock);
+
+	/* Copy the datum */
+	memcpy(valuePtr, DatumGetPointer(value), valueSize);
+	HnswPtrStore(base, element->value, valuePtr);
+
+	/* Create a lock for the element */
+	LWLockInitialize(&element->lock, hnsw_lock_tranche_id);
+
+	/* Insert tuple */
+	InsertTupleInMemory(buildstate, element);
+
+	/* Release flush lock */
+	LWLockRelease(flushLock);
+
+	return true;
 }

 /*
@@ -362,10 +585,8 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 			  bool *isnull, bool tupleIsAlive, void *state)
 {
 	HnswBuildState *buildstate = (HnswBuildState *) state;
+	HnswGraph  *graph = buildstate->graph;
 	MemoryContext oldCtx;
-	HnswElement element;
-	HnswElement dup = NULL;
-	bool		inserted;

 #if PG_VERSION_NUM < 130000
 	ItemPointer tid = &hup->t_self;
@@ -375,70 +596,80 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	if (isnull[0])
 		return;

-	if (buildstate->flushed || buildstate->memoryLeft <= 0 || list_length(buildstate->elements) == LIST_MAX_LENGTH)
-	{
-		if (!buildstate->flushed)
-		{
-			if (buildstate->memoryLeft <= 0)
-				ereport(NOTICE,
-						(errmsg("hnsw graph no longer fits into maintenance_work_mem after " INT64_FORMAT " tuples", (int64) buildstate->indtuples),
-						 errdetail("Building will take significantly more time."),
-						 errhint("Increase maintenance_work_mem to speed up builds.")));
-
-			FlushPages(buildstate);
-		}
-
-		oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
-
-		if (HnswInsertTuple(buildstate->index, values, isnull, tid, buildstate->heap, true))
-		{
-			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);
-		MemoryContextReset(buildstate->tmpCtx);
-
-		return;
-	}
-
-	/* Allocate necessary memory outside of memory context */
-	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
-
-	/* Use memory context since detoast can allocate */
+	/* Use memory context */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);

 	/* Insert tuple */
-	inserted = InsertTuple(index, values, element, buildstate, &dup, oldCtx);
+	if (InsertTuple(index, values, isnull, tid, buildstate))
+	{
+		/* Update progress */
+		SpinLockAcquire(&graph->lock);
+		UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++graph->indtuples);
+		SpinLockRelease(&graph->lock);
+	}

 	/* Reset memory context */
 	MemoryContextSwitchTo(oldCtx);
 	MemoryContextReset(buildstate->tmpCtx);
+}

-	/* Add outside memory context */
-	if (dup != NULL)
-	{
-		HnswAddHeapTid(dup, tid);
-		buildstate->memoryLeft -= sizeof(ItemPointerData);
-	}
+/*
+ * Initialize the graph
+ */
+static void
+InitGraph(HnswGraph * graph, char *base, long memoryTotal)
+{
+	HnswPtrStore(base, graph->head, (HnswElement) NULL);
+	HnswPtrStore(base, graph->entryPoint, (HnswElement) NULL);
+	graph->memoryUsed = 0;
+	graph->memoryTotal = memoryTotal;
+	graph->flushed = false;
+	graph->indtuples = 0;
+	SpinLockInit(&graph->lock);
+	LWLockInitialize(&graph->entryLock, hnsw_lock_tranche_id);
+	LWLockInitialize(&graph->allocatorLock, hnsw_lock_tranche_id);
+	LWLockInitialize(&graph->flushLock, hnsw_lock_tranche_id);
+}

-	/* Add to buildstate or free */
-	if (inserted)
-	{
-		buildstate->elements = lappend(buildstate->elements, element);
-		buildstate->memoryLeft -= HnswElementMemory(element, buildstate->m);
-	}
-	else
-		HnswFreeElement(element);
+/*
+ * Initialize an allocator
+ */
+static void
+InitAllocator(HnswAllocator * allocator, void *(*alloc) (Size size, void *state), void *state)
+{
+	allocator->alloc = alloc;
+	allocator->state = state;
+}
+
+/*
+ * Memory context allocator
+ */
+static void *
+HnswMemoryContextAlloc(Size size, void *state)
+{
+	HnswBuildState *buildstate = (HnswBuildState *) state;
+	void	   *chunk = MemoryContextAlloc(buildstate->graphCtx, size);
+
+#if PG_VERSION_NUM >= 130000
+	buildstate->graphData.memoryUsed = MemoryContextMemAllocated(buildstate->graphCtx, false);
+#else
+	buildstate->graphData.memoryUsed += MAXALIGN(size);
+#endif
+
+	return chunk;
+}
+
+/*
+ * Shared memory allocator
+ */
+static void *
+HnswSharedMemoryAlloc(Size size, void *state)
+{
+	HnswBuildState *buildstate = (HnswBuildState *) state;
+	void	   *chunk = buildstate->hnswarea + buildstate->graph->memoryUsed;
+
+	buildstate->graph->memoryUsed += MAXALIGN(size);
+	return chunk;
 }

 /*
@@ -474,22 +705,29 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	buildstate->collation = index->rd_indcollation[0];

-	buildstate->elements = NIL;
-	buildstate->entryPoint = NULL;
+	InitGraph(&buildstate->graphData, NULL, maintenance_work_mem * 1024L);
+	buildstate->graph = &buildstate->graphData;
 	buildstate->ml = HnswGetMl(buildstate->m);
 	buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
-	buildstate->memoryLeft = maintenance_work_mem * 1024L;
-	buildstate->flushed = false;

 	/* Reuse for each tuple */
 	buildstate->normvec = InitVector(buildstate->dimensions);

+	buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
+												   "Hnsw build graph context",
+#if PG_VERSION_NUM >= 150000
+												   1024 * 1024, 1024 * 1024,
+#endif
+												   1024 * 1024);
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
 											   "Hnsw build temporary context",
 											   ALLOCSET_DEFAULT_SIZES);

+	InitAllocator(&buildstate->allocator, &HnswMemoryContextAlloc, buildstate);
+
 	buildstate->hnswleader = NULL;
 	buildstate->hnswshared = NULL;
+	buildstate->hnswarea = NULL;
 }

 /*
@@ -499,6 +737,7 @@ static void
 FreeBuildState(HnswBuildState * buildstate)
 {
 	pfree(buildstate->normvec);
+	MemoryContextDelete(buildstate->graphCtx);
 	MemoryContextDelete(buildstate->tmpCtx);
 }

@@ -518,7 +757,8 @@ ParallelHeapScan(HnswBuildState * buildstate)
 		SpinLockAcquire(&hnswshared->mutex);
 		if (hnswshared->nparticipantsdone == nparticipanttuplesorts)
 		{
-			buildstate->indtuples = hnswshared->indtuples;
+			buildstate->graph = &hnswshared->graphData;
+			buildstate->hnswarea = buildstate->hnswleader->hnswarea;
 			reltuples = hnswshared->reltuples;
 			SpinLockRelease(&hnswshared->mutex);
 			break;
@@ -538,7 +778,7 @@ ParallelHeapScan(HnswBuildState * buildstate)
 * Perform a worker's portion of a parallel insert
 */
 static void
-HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool progress)
+HnswParallelScanAndInsert(Relation heapRel, Relation indexRel, HnswShared * hnswshared, char *hnswarea, bool progress)
 {
 	HnswBuildState buildstate;
 #if PG_VERSION_NUM >= 120000
@@ -550,22 +790,21 @@ HnswParallelScanAndInsert(HnswSpool * hnswspool, HnswShared * hnswshared, bool p
 	IndexInfo  *indexInfo;

 	/* Join parallel scan */
-	indexInfo = BuildIndexInfo(hnswspool->index);
+	indexInfo = BuildIndexInfo(indexRel);
 	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;
+	InitBuildState(&buildstate, heapRel, indexRel, indexInfo, MAIN_FORKNUM);
+	buildstate.graph = &hnswshared->graphData;
+	buildstate.hnswarea = hnswarea;
+	InitAllocator(&buildstate.allocator, &HnswSharedMemoryAlloc, &buildstate);
 #if PG_VERSION_NUM >= 120000
-	scan = table_beginscan_parallel(hnswspool->heap,
+	scan = table_beginscan_parallel(heapRel,
 									ParallelTableScanFromHnswShared(hnswshared));
-	reltuples = table_index_build_scan(hnswspool->heap, hnswspool->index, indexInfo,
+	reltuples = table_index_build_scan(heapRel, indexRel, indexInfo,
 									   true, progress, BuildCallback,
 									   (void *) &buildstate, scan);
 #else
-	scan = heap_beginscan_parallel(hnswspool->heap, &hnswshared->heapdesc);
-	reltuples = IndexBuildHeapScan(hnswspool->heap, hnswspool->index, indexInfo,
+	scan = heap_beginscan_parallel(heapRel, &hnswshared->heapdesc);
+	reltuples = IndexBuildHeapScan(heapRel, indexRel, indexInfo,
 								   true, BuildCallback,
 								   (void *) &buildstate, scan);
 #endif
@@ -595,8 +834,8 @@ void
 HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
 {
 	char	   *sharedquery;
-	HnswSpool  *hnswspool;
 	HnswShared *hnswshared;
+	char	   *hnswarea;
 	Relation	heapRel;
 	Relation	indexRel;
 	LOCKMODE	heapLockmode;
@@ -632,13 +871,10 @@ HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
 #endif
 	indexRel = index_open(hnswshared->indexrelid, indexLockmode);

-	/* Initialize worker's own spool */
-	hnswspool = (HnswSpool *) palloc0(sizeof(HnswSpool));
-	hnswspool->heap = heapRel;
-	hnswspool->index = indexRel;
+	hnswarea = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_AREA, false);

 	/* Perform inserts */
-	HnswParallelScanAndInsert(hnswspool, hnswshared, false);
+	HnswParallelScanAndInsert(heapRel, indexRel, hnswshared, hnswarea, false);

 	/* Close relations within worker */
 	index_close(indexRel, indexLockmode);
@@ -693,15 +929,9 @@ 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);
+	HnswParallelScanAndInsert(buildstate->heap, buildstate->index, hnswleader->hnswshared, hnswleader->hnswarea, true);
 }

 /*
@@ -711,10 +941,12 @@ static void
 HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 {
 	ParallelContext *pcxt;
-	int			scantuplesortstates;
 	Snapshot	snapshot;
 	Size		esthnswshared;
+	Size		esthnswarea;
+	Size		estother;
 	HnswShared *hnswshared;
+	char	   *hnswarea;
 	HnswLeader *hnswleader = (HnswLeader *) palloc0(sizeof(HnswLeader));
 	bool		leaderparticipates = true;
 	int			querylen;
@@ -732,8 +964,6 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	pcxt = CreateParallelContext("vector", "HnswParallelBuildMain", request, true);
 #endif

-	scantuplesortstates = leaderparticipates ? request + 1 : request;
-
 	/* Get snapshot for table scan */
 	if (!isconcurrent)
 		snapshot = SnapshotAny;
@@ -743,7 +973,17 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	/* Estimate size of workspaces */
 	esthnswshared = ParallelEstimateShared(buildstate->heap, snapshot);
 	shm_toc_estimate_chunk(&pcxt->estimator, esthnswshared);
-	shm_toc_estimate_keys(&pcxt->estimator, 1);
+
+	/* 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;
+	estother = 2 * 1024 * 1024;
+	if (esthnswarea > estother)
+		esthnswarea -= estother;
+
+	shm_toc_estimate_chunk(&pcxt->estimator, esthnswarea);
+	shm_toc_estimate_keys(&pcxt->estimator, 2);

 	/* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
 	if (debug_query_string)
@@ -774,13 +1014,11 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	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),
@@ -789,7 +1027,12 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	heap_parallelscan_initialize(&hnswshared->heapdesc, buildstate->heap, snapshot);
 #endif

+	hnswarea = (char *) shm_toc_allocate(pcxt->toc, esthnswarea);
+	/* Report less than allocated so never fails */
+	InitGraph(&hnswshared->graphData, hnswarea, esthnswarea - 1024 * 1024);
+
 	shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
+	shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_AREA, hnswarea);

 	/* Store query string for workers */
 	if (debug_query_string)
@@ -809,6 +1052,7 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 		hnswleader->nparticipanttuplesorts++;
 	hnswleader->hnswshared = hnswshared;
 	hnswleader->snapshot = snapshot;
+	hnswleader->hnswarea = hnswarea;

 	/* If no workers were successfully launched, back out (do serial build) */
 	if (pcxt->nworkers_launched == 0)
@@ -831,6 +1075,27 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
 	WaitForParallelWorkersToAttach(pcxt);
 }

+/*
+ * Compute parallel workers
+ */
+static int
+ComputeParallelWorkers(Relation heap, Relation index)
+{
+	int			parallel_workers;
+
+	/* Make sure it's safe to use parallel workers */
+	parallel_workers = plan_create_index_workers(RelationGetRelid(heap), RelationGetRelid(index));
+	if (parallel_workers == 0)
+		return 0;
+
+	/* Use parallel_workers storage parameter on table if set */
+	parallel_workers = RelationGetParallelWorkers(heap, -1);
+	if (parallel_workers != -1)
+		return Min(parallel_workers, max_parallel_maintenance_workers);
+
+	return max_parallel_maintenance_workers;
+}
+
 /*
 * Build graph
 */
@@ -842,30 +1107,35 @@ BuildGraph(HnswBuildState * buildstate, ForkNumber forkNum)
 	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));
+	if (buildstate->heap != NULL)
+		parallel_workers = ComputeParallelWorkers(buildstate->heap, 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 graph */
+	if (buildstate->heap != NULL)
+	{
+		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,
+														   true, true, BuildCallback, (void *) buildstate, NULL);
+#else
+			buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
+													   true, BuildCallback, (void *) buildstate, NULL);
+#endif
+		}
+
+		buildstate->indtuples = buildstate->graph->indtuples;
 	}

-	/* 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,
-													   true, true, BuildCallback, (void *) buildstate, NULL);
-#else
-		buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
-												   true, BuildCallback, (void *) buildstate, NULL);
-#endif
-	}
+	/* Flush pages */
+	if (!buildstate->graph->flushed)
+		FlushPages(buildstate);

 	/* End parallel build */
 	if (buildstate->hnswleader)
@@ -895,13 +1165,13 @@ static void
 BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
 		   HnswBuildState * buildstate, ForkNumber forkNum)
 {
+#ifdef HNSW_MEMORY
+	SeedRandom(42);
+#endif
+
 	InitBuildState(buildstate, heap, index, indexInfo, forkNum);

-	if (buildstate->heap != NULL)
-		BuildGraph(buildstate, forkNum);
-
-	if (!buildstate->flushed)
-		FlushPages(buildstate);
+	BuildGraph(buildstate, forkNum);

 	if (RelationNeedsWAL(index))
 		log_newpage_range(index, forkNum, 0, RelationGetNumberOfBlocks(index), true);
--- a/src/hnswinsert.c
+++ b/src/hnswinsert.c
@@ -5,6 +5,7 @@
 #include "hnsw.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"

 /*
@@ -115,7 +116,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
 * Add to element and neighbor pages
 */
 static void
-WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage, bool building)
+AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage, bool building)
 {
 	Buffer		buf;
 	Page		page;
@@ -133,9 +134,10 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	OffsetNumber freeOffno = InvalidOffsetNumber;
 	OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
 	BlockNumber newInsertPage = InvalidBlockNumber;
+	char	   *base = NULL;

 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(DatumGetPointer(e->value)));
+	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;
@@ -143,11 +145,11 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag

 	/* Prepare element tuple */
 	etup = palloc0(etupSize);
-	HnswSetElementTuple(etup, e);
+	HnswSetElementTuple(base, etup, e);

 	/* Prepare neighbor tuple */
 	ntup = palloc0(ntupSize);
-	HnswSetNeighborTuple(ntup, e, m);
+	HnswSetNeighborTuple(base, ntup, e, m);

 	/* Find a page (or two if needed) to insert the tuples */
 	for (;;)
@@ -219,7 +221,9 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 			HnswInsertAppendPage(index, &newbuf, &newpage, state, page, building);

 			/* Commit */
-			if (!building)
+			if (building)
+				MarkBufferDirty(buf);
+			else
 				GenericXLogFinish(state);

 			/* Unlock previous buffer */
@@ -294,7 +298,13 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	}

 	/* Commit */
-	if (!building)
+	if (building)
+	{
+		MarkBufferDirty(buf);
+		if (nbuf != buf)
+			MarkBufferDirty(nbuf);
+	}
+	else
 		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 	if (nbuf != buf)
@@ -329,12 +339,14 @@ ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
 * Update neighbors
 */
 void
-HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
+HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
 {
+	char	   *base = NULL;
+
 	for (int lc = e->level; lc >= 0; lc--)
 	{
 		int			lm = HnswGetLayerM(m, lc);
-		HnswNeighborArray *neighbors = &e->neighbors[lc];
+		HnswNeighborArray *neighbors = HnswGetNeighbors(base, e, lc);

 		for (int i = 0; i < neighbors->length; i++)
 		{
@@ -347,11 +359,12 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 			Size		ntupSize;
 			int			idx = -1;
 			int			startIdx;
-			OffsetNumber offno = hc->element->neighborOffno;
+			HnswElement neighborElement = HnswPtrAccess(base, hc->element);
+			OffsetNumber offno = neighborElement->neighborOffno;

 			/* Get latest neighbors since they may have changed */
 			/* Do not lock yet since selecting neighbors can take time */
-			HnswLoadNeighbors(hc->element, index, m);
+			HnswLoadNeighbors(neighborElement, index, m);

 			/*
 			 * Could improve performance for vacuuming by checking neighbors
@@ -361,14 +374,14 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 			 */

 			/* Select neighbors */
-			HnswUpdateConnection(e, hc, lm, lc, &idx, index, procinfo, collation);
+			HnswUpdateConnection(NULL, e, hc, lm, lc, &idx, index, procinfo, collation);

 			/* New element was not selected as a neighbor */
 			if (idx == -1)
 				continue;

 			/* Register page */
-			buf = ReadBuffer(index, hc->element->neighborPage);
+			buf = ReadBuffer(index, neighborElement->neighborPage);
 			LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
 			if (building)
 			{
@@ -387,7 +400,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 			ntupSize = ItemIdGetLength(itemid);

 			/* Calculate index for update */
-			startIdx = (hc->element->level - lc) * m;
+			startIdx = (neighborElement->level - lc) * m;

 			/* Check for existing connection */
 			if (checkExisting && ConnectionExists(e, ntup, startIdx, lm))
@@ -421,7 +434,9 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 					elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));

 				/* Commit */
-				if (!building)
+				if (building)
+					MarkBufferDirty(buf);
+				else
 					GenericXLogFinish(state);
 			}
 			else if (!building)
@@ -436,7 +451,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
 * Add a heap TID to an existing element
 */
 static bool
-HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool building)
+AddDuplicateOnDisk(Relation index, HnswElement element, HnswElement dup, bool building)
 {
 	Buffer		buf;
 	Page		page;
@@ -480,14 +495,16 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool buil
 	}

 	/* Add heap TID */
-	etup->heaptids[i] = *((ItemPointer) linitial(element->heaptids));
+	etup->heaptids[i] = element->heaptids[0];

 	/* Overwrite tuple */
 	if (!PageIndexTupleOverwrite(page, dup->offno, (Item) etup, etupSize))
 		elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));

 	/* Commit */
-	if (!building)
+	if (building)
+		MarkBufferDirty(buf);
+	else
 		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);

@@ -495,31 +512,55 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup, bool buil
 }

 /*
- * Write changes to disk
+ * Find duplicate element
+ */
+static bool
+FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
+{
+	char	   *base = NULL;
+	HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
+	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 (!datumIsEqual(value, neighborValue, false, -1))
+			return false;
+
+		if (AddDuplicateOnDisk(index, element, neighborElement, building))
+			return true;
+	}
+
+	return false;
+}
+
+/*
+ * Update graph on disk
 */
 static void
-WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement dup, HnswElement entryPoint, bool building)
+UpdateGraphOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
 {
 	BlockNumber newInsertPage = InvalidBlockNumber;

-	/* Try to add to existing page */
-	if (dup != NULL)
-	{
-		if (HnswAddDuplicate(index, element, dup, building))
-			return;
-	}
+	/* Look for duplicate */
+	if (FindDuplicateOnDisk(index, element, building))
+		return;

-	/* Write element and neighbor tuples */
-	WriteNewElementPages(index, element, m, GetInsertPage(index), &newInsertPage, building);
+	/* Add element */
+	AddElementOnDisk(index, element, m, GetInsertPage(index), &newInsertPage, building);

 	/* Update insert page if needed */
 	if (BlockNumberIsValid(newInsertPage))
 		HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);

 	/* Update neighbors */
-	HnswUpdateNeighborPages(index, procinfo, collation, element, m, false, building);
+	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, false, building);

-	/* Update metapage if needed */
+	/* Update entry point if needed */
 	if (entryPoint == NULL || element->level > entryPoint->level)
 		HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, building);
 }
@@ -528,29 +569,16 @@ WriteElement(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement elem
 * Insert a tuple into the index
 */
 bool
-HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building)
+HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel, bool building)
 {
-	Datum		value;
-	FmgrInfo   *normprocinfo;
 	HnswElement entryPoint;
 	HnswElement element;
 	int			m;
 	int			efConstruction = HnswGetEfConstruction(index);
 	FmgrInfo   *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	Oid			collation = index->rd_indcollation[0];
-	HnswElement dup;
 	LOCKMODE	lockmode = ShareLock;
-
-	/* Detoast once for all calls */
-	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
-
-	/* Normalize if needed */
-	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
-	if (normprocinfo != NULL)
-	{
-		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
-			return false;
-	}
+	char	   *base = NULL;

 	/*
 	 * Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -563,8 +591,8 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	HnswGetMetaPageInfo(index, &m, &entryPoint);

 	/* Create an element */
-	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->value = value;
+	element = HnswInitElement(base, heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
+	HnswPtrStore(base, element->value, DatumGetPointer(value));

 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)
@@ -580,14 +608,11 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 		entryPoint = HnswGetEntryPoint(index);
 	}

-	/* Insert element in graph */
-	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, false);
+	/* Find neighbors for element */
+	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, false);

-	/* Look for duplicate */
-	dup = HnswFindDuplicate(element);
-
-	/* Write to disk */
-	WriteElement(index, procinfo, collation, element, m, efConstruction, dup, entryPoint, building);
+	/* Update graph on disk */
+	UpdateGraphOnDisk(index, procinfo, collation, element, m, efConstruction, entryPoint, building);

 	/* Release lock */
 	UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
@@ -595,6 +620,30 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	return true;
 }

+/*
+ * Insert a tuple into the index
+ */
+static void
+HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
+{
+	Datum		value;
+	FmgrInfo   *normprocinfo;
+	Oid			collation = index->rd_indcollation[0];
+
+	/* Detoast once for all calls */
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+
+	/* Normalize if needed */
+	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	if (normprocinfo != NULL)
+	{
+		if (!HnswNormValue(normprocinfo, collation, &value, NULL))
+			return;
+	}
+
+	HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, heapRel, false);
+}
+
 /*
 * Insert a tuple into the index
 */
@@ -621,7 +670,7 @@ hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
 	oldCtx = MemoryContextSwitchTo(insertCtx);

 	/* Insert tuple */
-	HnswInsertTuple(index, values, isnull, heap_tid, heap, false);
+	HnswInsertTuple(index, values, isnull, heap_tid, heap);

 	/* Delete memory context */
 	MemoryContextSwitchTo(oldCtx);
--- a/src/hnswscan.c
+++ b/src/hnswscan.c
@@ -21,6 +21,7 @@ GetScanItems(IndexScanDesc scan, Datum q)
 	List	   *w;
 	int			m;
 	HnswElement entryPoint;
+	char	   *base = NULL;

 	/* Get m and entry point */
 	HnswGetMetaPageInfo(index, &m, &entryPoint);
@@ -28,15 +29,15 @@ GetScanItems(IndexScanDesc scan, Datum q)
 	if (entryPoint == NULL)
 		return NIL;

-	ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, procinfo, collation, false));

 	for (int lc = entryPoint->level; lc >= 1; lc--)
 	{
-		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
 		ep = w;
 	}

-	return HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
+	return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
 }

 /*
@@ -184,19 +185,19 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)

 	while (list_length(so->w) > 0)
 	{
+		char	   *base = NULL;
 		HnswCandidate *hc = llast(so->w);
+		HnswElement element = HnswPtrAccess(base, hc->element);
 		ItemPointer heaptid;

 		/* Move to next element if no valid heap TIDs */
-		if (list_length(hc->element->heaptids) == 0)
+		if (element->heaptidsLength == 0)
 		{
 			so->w = list_delete_last(so->w);
 			continue;
 		}

-		heaptid = llast(hc->element->heaptids);
-
-		hc->element->heaptids = list_delete_last(hc->element->heaptids);
+		heaptid = &element->heaptids[--element->heaptidsLength];

 		MemoryContextSwitchTo(oldCtx);

--- a/src/hnswutils.c
+++ b/src/hnswutils.c
@@ -84,9 +84,40 @@ hash_pointer(uintptr_t ptr)
 #define SH_DEFINE
 #include "lib/simplehash.h"

+/* Needed to include simplehash.h again */
+#if PG_VERSION_NUM < 120000
+#undef SH_EQUAL
+#undef sh_log2
+#undef sh_pow2
+#define sh_log2 offsethash_sh_log2
+#define sh_pow2 offsethash_sh_pow2
+#endif
+
+/* Offset hash table */
+static uint32
+hash_offset(Size offset)
+{
+#if SIZEOF_SIZE_T == 8
+	return murmurhash64((uint64) offset);
+#else
+	return murmurhash32((uint32) offset);
+#endif
+}
+
+#define SH_PREFIX		offsethash
+#define SH_ELEMENT_TYPE	OffsetHashEntry
+#define SH_KEY_TYPE		Size
+#define	SH_KEY			offset
+#define SH_HASH_KEY(tb, key)	hash_offset(key)
+#define SH_EQUAL(tb, a, b)		(a == b)
+#define	SH_SCOPE		extern
+#define SH_DEFINE
+#include "lib/simplehash.h"
+
 typedef union
 {
 	pointerhash_hash *pointers;
+	offsethash_hash *offsets;
 	tidhash_hash *tids;
 }			visited_hash;

@@ -188,42 +219,46 @@ HnswInitPage(Buffer buf, Page page)
 * Allocate neighbors
 */
 void
-HnswInitNeighbors(HnswElement element, int m)
+HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * allocator)
 {
 	int			level = element->level;

-	element->neighbors = palloc(sizeof(HnswNeighborArray) * (level + 1));
+	HnswNeighborArrayPtr *neighborList = (HnswNeighborArrayPtr *) HnswAlloc(allocator, sizeof(HnswNeighborArrayPtr) * (level + 1));
+
+	HnswPtrStore(base, element->neighbors, neighborList);

 	for (int lc = 0; lc <= level; lc++)
 	{
 		HnswNeighborArray *a;
 		int			lm = HnswGetLayerM(m, lc);

-		a = &element->neighbors[lc];
+		HnswPtrStore(base, neighborList[lc], (HnswNeighborArray *) HnswAlloc(allocator, offsetof(HnswNeighborArray, items) + sizeof(HnswCandidate) * lm));
+
+		a = HnswGetNeighbors(base, element, lc);
 		a->length = 0;
-		a->items = palloc(sizeof(HnswCandidate) * lm);
 		a->closerSet = false;
 	}
 }

 /*
- * Free neighbors
+ * Allocate memory from the allocator
 */
-static void
-HnswFreeNeighbors(HnswElement element)
+void *
+HnswAlloc(HnswAllocator * allocator, Size size)
 {
-	for (int lc = 0; lc <= element->level; lc++)
-		pfree(element->neighbors[lc].items);
-	pfree(element->neighbors);
+	if (allocator)
+		return (*(allocator)->alloc) (size, (allocator)->state);
+
+	return palloc(size);
 }

 /*
 * Allocate an element
 */
 HnswElement
-HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
+HnswInitElement(char *base, ItemPointer heaptid, int m, double ml, int maxLevel, HnswAllocator * allocator)
 {
-	HnswElement element = palloc(sizeof(HnswElementData));
+	HnswElement element = HnswAlloc(allocator, sizeof(HnswElementData));

 	int			level = (int) (-log(RandomDouble()) * ml);

@@ -231,42 +266,26 @@ HnswInitElement(ItemPointer heaptid, int m, double ml, int maxLevel)
 	if (level > maxLevel)
 		level = maxLevel;

-	element->heaptids = NIL;
+	element->heaptidsLength = 0;
 	HnswAddHeapTid(element, heaptid);

 	element->level = level;
 	element->deleted = 0;

-	HnswInitNeighbors(element, m);
+	HnswInitNeighbors(base, element, m, allocator);

-	element->value = PointerGetDatum(NULL);
+	HnswPtrStore(base, element->value, (Pointer) NULL);

 	return element;
 }

-/*
- * Free an element
- */
-void
-HnswFreeElement(HnswElement element)
-{
-	HnswFreeNeighbors(element);
-	list_free_deep(element->heaptids);
-	if (DatumGetPointer(element->value))
-		pfree(DatumGetPointer(element->value));
-	pfree(element);
-}
-
 /*
 * Add a heap TID to an element
 */
 void
 HnswAddHeapTid(HnswElement element, ItemPointer heaptid)
 {
-	ItemPointer copy = palloc(sizeof(ItemPointerData));
-
-	ItemPointerCopy(heaptid, copy);
-	element->heaptids = lappend(element->heaptids, copy);
+	element->heaptids[element->heaptidsLength++] = *heaptid;
 }

 /*
@@ -276,11 +295,12 @@ HnswElement
 HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 {
 	HnswElement element = palloc(sizeof(HnswElementData));
+	char	   *base = NULL;

 	element->blkno = blkno;
 	element->offno = offno;
-	element->neighbors = NULL;
-	element->value = PointerGetDatum(NULL);
+	HnswPtrStore(base, element->neighbors, (HnswNeighborArrayPtr *) NULL);
+	HnswPtrStore(base, element->value, (Pointer) NULL);
 	return element;
 }

@@ -379,7 +399,9 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc

 	HnswUpdateMetaPageInfo(page, updateEntry, entryPoint, insertPage);

-	if (!building)
+	if (building)
+		MarkBufferDirty(buf);
+	else
 		GenericXLogFinish(state);
 	UnlockReleaseBuffer(buf);
 }
@@ -388,26 +410,28 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
 * Set element tuple, except for neighbor info
 */
 void
-HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
+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;
 	for (int i = 0; i < HNSW_HEAPTIDS; i++)
 	{
-		if (i < list_length(element->heaptids))
-			etup->heaptids[i] = *((ItemPointer) list_nth(element->heaptids, i));
+		if (i < element->heaptidsLength)
+			etup->heaptids[i] = element->heaptids[i];
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->data, DatumGetPointer(element->value), VARSIZE_ANY(DatumGetPointer(element->value)));
+	memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
 }

 /*
 * Set neighbor tuple
 */
 void
-HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
+HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m)
 {
 	int			idx = 0;

@@ -415,7 +439,7 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)

 	for (int lc = e->level; lc >= 0; lc--)
 	{
-		HnswNeighborArray *neighbors = &e->neighbors[lc];
+		HnswNeighborArray *neighbors = HnswGetNeighbors(base, e, lc);
 		int			lm = HnswGetLayerM(m, lc);

 		for (int i = 0; i < lm; i++)
@@ -425,8 +449,9 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
 			if (i < neighbors->length)
 			{
 				HnswCandidate *hc = &neighbors->items[i];
+				HnswElement hce = HnswPtrAccess(base, hc->element);

-				ItemPointerSet(indextid, hc->element->blkno, hc->element->offno);
+				ItemPointerSet(indextid, hce->blkno, hce->offno);
 			}
 			else
 				ItemPointerSetInvalid(indextid);
@@ -442,12 +467,14 @@ HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m)
 static void
 LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
 {
+	char	   *base = NULL;
+
 	HnswNeighborTuple ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
 	int			neighborCount = (element->level + 2) * m;

 	Assert(HnswIsNeighborTuple(ntup));

-	HnswInitNeighbors(element, m);
+	HnswInitNeighbors(base, element, m, NULL);

 	/* Ensure expected neighbors */
 	if (ntup->count != neighborCount)
@@ -473,9 +500,9 @@ LoadNeighborsFromPage(HnswElement element, Relation index, Page page, int m)
 		if (level < 0)
 			level = 0;

-		neighbors = &element->neighbors[level];
+		neighbors = HnswGetNeighbors(base, element, level);
 		hc = &neighbors->items[neighbors->length++];
-		hc->element = e;
+		HnswPtrStore(base, hc->element, e);
 	}
 }

@@ -507,7 +534,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	element->deleted = etup->deleted;
 	element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
 	element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
-	element->heaptids = NIL;
+	element->heaptidsLength = 0;

 	if (loadHeaptids)
 	{
@@ -522,7 +549,12 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 	}

 	if (loadVec)
-		element->value = datumCopy(PointerGetDatum(&etup->data), false, -1);
+	{
+		char	   *base = NULL;
+		Datum		value = datumCopy(PointerGetDatum(&etup->data), false, -1);
+
+		HnswPtrStore(base, element->value, DatumGetPointer(value));
+	}
 }

 /*
@@ -558,24 +590,27 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 * Get the distance for a candidate
 */
 static float
-GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
+GetCandidateDistance(char *base, HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
 {
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
+	HnswElement hce = HnswPtrAccess(base, hc->element);
+	Datum		value = HnswGetValue(base, hce);
+
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, value));
 }

 /*
 * Create a candidate for the entry point
 */
 HnswCandidate *
-HnswEntryCandidate(HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
+HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
 {
 	HnswCandidate *hc = palloc(sizeof(HnswCandidate));

-	hc->element = entryPoint;
+	HnswPtrStore(base, hc->element, entryPoint);
 	if (index == NULL)
-		hc->distance = GetCandidateDistance(hc, q, procinfo, collation);
+		hc->distance = GetCandidateDistance(base, hc, q, procinfo, collation);
 	else
-		HnswLoadElement(hc->element, &hc->distance, &q, index, procinfo, collation, loadVec);
+		HnswLoadElement(entryPoint, &hc->distance, &q, index, procinfo, collation, loadVec);
 	return hc;
 }

@@ -621,48 +656,97 @@ CreatePairingHeapNode(HnswCandidate * c)
 	return node;
 }

+/*
+ * Init visited
+ */
+static inline void
+InitVisited(char *base, visited_hash * v, Relation index, int ef, int m)
+{
+	if (index != NULL)
+		v->tids = tidhash_create(CurrentMemoryContext, ef * m * 2, NULL);
+	else if (base != NULL)
+		v->offsets = offsethash_create(CurrentMemoryContext, ef * m * 2, NULL);
+	else
+		v->pointers = pointerhash_create(CurrentMemoryContext, ef * m * 2, NULL);
+}
+
 /*
 * Add to visited
 */
 static inline void
-AddToVisited(visited_hash v, HnswCandidate * hc, Relation index, bool *found)
+AddToVisited(char *base, visited_hash * v, HnswCandidate * hc, Relation index, bool *found)
 {
-	if (index == NULL)
+	if (index != NULL)
+	{
+		HnswElement element = HnswPtrAccess(base, hc->element);
+		ItemPointerData indextid;
+
+		ItemPointerSet(&indextid, element->blkno, element->offno);
+		tidhash_insert(v->tids, indextid, found);
+	}
+	else if (base != NULL)
 	{
 #if PG_VERSION_NUM >= 130000
-		pointerhash_insert_hash(v.pointers, (uintptr_t) hc->element, hc->element->hash, found);
+		HnswElement element = HnswPtrAccess(base, hc->element);
+
+		offsethash_insert_hash(v->offsets, HnswPtrOffset(hc->element), element->hash, found);
 #else
-		pointerhash_insert(v.pointers, (uintptr_t) hc->element, found);
+		offsethash_insert(v->offsets, HnswPtrOffset(hc->element), found);
 #endif
 	}
 	else
 	{
-		ItemPointerData indextid;
+#if PG_VERSION_NUM >= 130000
+		HnswElement element = HnswPtrAccess(base, hc->element);

-		ItemPointerSet(&indextid, hc->element->blkno, hc->element->offno);
-		tidhash_insert(v.tids, indextid, found);
+		pointerhash_insert_hash(v->pointers, (uintptr_t) HnswPtrPointer(hc->element), element->hash, found);
+#else
+		pointerhash_insert(v->pointers, (uintptr_t) HnswPtrPointer(hc->element), found);
+#endif
 	}
 }

+/*
+ * Count element towards ef
+ */
+static inline bool
+CountElement(char *base, HnswElement skipElement, HnswCandidate * hc)
+{
+	HnswElement e;
+
+	if (skipElement == NULL)
+		return true;
+
+	/* Ensure does not access heaptidsLength during in-memory build */
+	pg_memory_barrier();
+
+	e = HnswPtrAccess(base, hc->element);
+	return e->heaptidsLength != 0;
+}
+
 /*
 * Algorithm 2 from paper
 */
 List *
-HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
+HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
 {
-	ListCell   *lc2;
-
 	List	   *w = NIL;
 	pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
 	pairingheap *W = pairingheap_allocate(CompareFurthestCandidates, NULL);
 	int			wlen = 0;
 	visited_hash v;
+	ListCell   *lc2;
+	HnswNeighborArray *neighborhoodData = NULL;
+	Size		neighborhoodSize;

-	/* Create hash table */
+	InitVisited(base, &v, index, ef, m);
+
+	/* Create local memory for neighborhood if needed */
 	if (index == NULL)
-		v.pointers = pointerhash_create(CurrentMemoryContext, ef * m * 2, NULL);
-	else
-		v.tids = tidhash_create(CurrentMemoryContext, ef * m * 2, NULL);
+	{
+		neighborhoodSize = offsetof(HnswNeighborArray, items) + sizeof(HnswCandidate) * HnswGetLayerM(m, lc);
+		neighborhoodData = palloc(neighborhoodSize);
+	}

 	/* Add entry points to v, C, and W */
 	foreach(lc2, ep)
@@ -670,7 +754,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 		HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
 		bool		found;

-		AddToVisited(v, hc, index, &found);
+		AddToVisited(base, &v, hc, index, &found);

 		pairingheap_add(C, &(CreatePairingHeapNode(hc)->ph_node));
 		pairingheap_add(W, &(CreatePairingHeapNode(hc)->ph_node));
@@ -680,7 +764,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 		 * would be ideal to do this for inserts as well, but this could
 		 * affect insert performance.
 		 */
-		if (skipElement == NULL || list_length(hc->element->heaptids) != 0)
+		if (CountElement(base, skipElement, hc))
 			wlen++;
 	}

@@ -689,38 +773,51 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 		HnswNeighborArray *neighborhood;
 		HnswCandidate *c = ((HnswPairingHeapNode *) pairingheap_remove_first(C))->inner;
 		HnswCandidate *f = ((HnswPairingHeapNode *) pairingheap_first(W))->inner;
+		HnswElement cElement;

 		if (c->distance > f->distance)
 			break;

-		if (c->element->neighbors == NULL)
-			HnswLoadNeighbors(c->element, index, m);
+		cElement = HnswPtrAccess(base, c->element);
+
+		if (HnswPtrIsNull(base, cElement->neighbors))
+			HnswLoadNeighbors(cElement, index, m);

 		/* Get the neighborhood at layer lc */
-		neighborhood = &c->element->neighbors[lc];
+		neighborhood = HnswGetNeighbors(base, cElement, lc);
+
+		/* Copy neighborhood to local memory if needed */
+		if (index == NULL)
+		{
+			LWLockAcquire(&cElement->lock, LW_SHARED);
+			memcpy(neighborhoodData, neighborhood, neighborhoodSize);
+			LWLockRelease(&cElement->lock);
+			neighborhood = neighborhoodData;
+		}

 		for (int i = 0; i < neighborhood->length; i++)
 		{
 			HnswCandidate *e = &neighborhood->items[i];
 			bool		visited;

-			AddToVisited(v, e, index, &visited);
+			AddToVisited(base, &v, e, index, &visited);

 			if (!visited)
 			{
 				float		eDistance;
+				HnswElement eElement = HnswPtrAccess(base, e->element);

 				f = ((HnswPairingHeapNode *) pairingheap_first(W))->inner;

 				if (index == NULL)
-					eDistance = GetCandidateDistance(e, q, procinfo, collation);
+					eDistance = GetCandidateDistance(base, e, q, procinfo, collation);
 				else
-					HnswLoadElement(e->element, &eDistance, &q, index, procinfo, collation, inserting);
+					HnswLoadElement(eElement, &eDistance, &q, index, procinfo, collation, inserting);

-				Assert(!e->element->deleted);
+				Assert(!eElement->deleted);

 				/* Make robust to issues */
-				if (e->element->level < lc)
+				if (eElement->level < lc)
 					continue;

 				if (eDistance < f->distance || wlen < ef)
@@ -728,7 +825,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 					/* Copy e */
 					HnswCandidate *ec = palloc(sizeof(HnswCandidate));

-					ec->element = e->element;
+					HnswPtrStore(base, ec->element, eElement);
 					ec->distance = eDistance;

 					pairingheap_add(C, &(CreatePairingHeapNode(ec)->ph_node));
@@ -739,7 +836,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 					 * vacuuming. It would be ideal to do this for inserts as
 					 * well, but this could affect insert performance.
 					 */
-					if (skipElement == NULL || list_length(e->element->heaptids) != 0)
+					if (CountElement(base, skipElement, e))
 					{
 						wlen++;

@@ -764,7 +861,7 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 }

 /*
- * Compare candidate distances
+ * Compare candidate distances with pointer tie-breaker
 */
 static int
 #if PG_VERSION_NUM >= 130000
@@ -782,10 +879,38 @@ CompareCandidateDistances(const void *a, const void *b)
 	if (hca->distance > hcb->distance)
 		return -1;

-	if (hca->element < hcb->element)
+	if (HnswPtrPointer(hca->element) < HnswPtrPointer(hcb->element))
 		return 1;

-	if (hca->element > hcb->element)
+	if (HnswPtrPointer(hca->element) > HnswPtrPointer(hcb->element))
+		return -1;
+
+	return 0;
+}
+
+/*
+ * Compare candidate distances with offset tie-breaker
+ */
+static int
+#if PG_VERSION_NUM >= 130000
+CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
+#else
+CompareCandidateDistancesOffset(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 (HnswPtrOffset(hca->element) < HnswPtrOffset(hcb->element))
+		return 1;
+
+	if (HnswPtrOffset(hca->element) > HnswPtrOffset(hcb->element))
 		return -1;

 	return 0;
@@ -795,46 +920,58 @@ CompareCandidateDistances(const void *a, const void *b)
 * Calculate the distance between elements
 */
 static float
-HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid collation)
+HnswGetDistance(char *base, HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid collation)
 {
+	Datum		aValue;
+	Datum		bValue;
+
 	/* Look for cached distance */
-	if (a->neighbors != NULL)
+	if (!HnswPtrIsNull(base, a->neighbors))
 	{
-		Assert(a->level >= lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(base, a, lc);

-		for (int i = 0; i < a->neighbors[lc].length; i++)
+		for (int i = 0; i < neighbors->length; i++)
 		{
-			if (a->neighbors[lc].items[i].element == b)
-				return a->neighbors[lc].items[i].distance;
+			HnswElement element = HnswPtrAccess(base, neighbors->items[i].element);
+
+			if (element == b)
+				return neighbors->items[i].distance;
 		}
 	}

-	if (b->neighbors != NULL)
+	if (!HnswPtrIsNull(base, b->neighbors))
 	{
-		Assert(b->level >= lc);
+		HnswNeighborArray *neighbors = HnswGetNeighbors(base, b, lc);

-		for (int i = 0; i < b->neighbors[lc].length; i++)
+		for (int i = 0; i < neighbors->length; i++)
 		{
-			if (b->neighbors[lc].items[i].element == a)
-				return b->neighbors[lc].items[i].distance;
+			HnswElement element = HnswPtrAccess(base, neighbors->items[i].element);
+
+			if (element == a)
+				return neighbors->items[i].distance;
 		}
 	}

-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
+	aValue = HnswGetValue(base, a);
+	bValue = HnswGetValue(base, b);
+
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, aValue, bValue));
 }

 /*
 * Check if an element is closer to q than any element from R
 */
 static bool
-CheckElementCloser(HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid collation)
+CheckElementCloser(char *base, HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid collation)
 {
+	HnswElement eElement = HnswPtrAccess(base, e->element);
 	ListCell   *lc2;

 	foreach(lc2, r)
 	{
 		HnswCandidate *ri = lfirst(lc2);
-		float		distance = HnswGetDistance(e->element, ri->element, lc, procinfo, collation);
+		HnswElement riElement = HnswPtrAccess(base, ri->element);
+		float		distance = HnswGetDistance(base, eElement, riElement, lc, procinfo, collation);

 		if (distance <= e->distance)
 			return false;
@@ -847,12 +984,13 @@ 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, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
+SelectNeighbors(char *base, 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;
+	HnswNeighborArray *neighbors = HnswGetNeighbors(base, e2, lc);
+	bool		mustCalculate = !neighbors->closerSet;
 	List	   *added = NIL;
 	bool		removedAny = false;

@@ -863,7 +1001,12 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE

 	/* Ensure order of candidates is deterministic for closer caching */
 	if (sortCandidates)
-		list_sort(w, CompareCandidateDistances);
+	{
+		if (base == NULL)
+			list_sort(w, CompareCandidateDistances);
+		else
+			list_sort(w, CompareCandidateDistancesOffset);
+	}

 	while (list_length(w) > 0 && list_length(r) < m)
 	{
@@ -874,7 +1017,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE

 		/* Use previous state of r and wd to skip work when possible */
 		if (mustCalculate)
-			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+			e->closer = CheckElementCloser(base, e, r, lc, procinfo, collation);
 		else if (list_length(added) > 0)
 		{
 			/*
@@ -883,7 +1026,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 			 */
 			if (e->closer)
 			{
-				e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
+				e->closer = CheckElementCloser(base, e, added, lc, procinfo, collation);

 				if (!e->closer)
 					removedAny = true;
@@ -896,7 +1039,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 				 */
 				if (removedAny)
 				{
-					e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+					e->closer = CheckElementCloser(base, e, r, lc, procinfo, collation);
 					if (e->closer)
 						added = lappend(added, e);
 				}
@@ -904,7 +1047,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 		}
 		else if (e == newCandidate)
 		{
-			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+			e->closer = CheckElementCloser(base, e, r, lc, procinfo, collation);
 			if (e->closer)
 				added = lappend(added, e);
 		}
@@ -916,7 +1059,7 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 	}

 	/* Cached value can only be used in future if sorted deterministically */
-	e2->neighbors[lc].closerSet = sortCandidates;
+	neighbors->closerSet = sortCandidates;

 	/* Keep pruned connections */
 	while (!pairingheap_is_empty(wd) && list_length(r) < m)
@@ -934,38 +1077,14 @@ SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswE
 	return r;
 }

-/*
- * Find duplicate element
- */
-HnswElement
-HnswFindDuplicate(HnswElement e)
-{
-	HnswNeighborArray *neighbors = &e->neighbors[0];
-
-	for (int i = 0; i < neighbors->length; i++)
-	{
-		HnswCandidate *neighbor = &neighbors->items[i];
-
-		/* Exit early since ordered by distance */
-		if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
-			break;
-
-		/* Check for space */
-		if (list_length(neighbor->element->heaptids) < HNSW_HEAPTIDS)
-			return neighbor->element;
-	}
-
-	return NULL;
-}
-
 /*
 * Add connections
 */
 static void
-AddConnections(HnswElement element, List *neighbors, int m, int lc)
+AddConnections(char *base, HnswElement element, List *neighbors, int m, int lc)
 {
 	ListCell   *lc2;
-	HnswNeighborArray *a = &element->neighbors[lc];
+	HnswNeighborArray *a = HnswGetNeighbors(base, element, lc);

 	foreach(lc2, neighbors)
 		a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
@@ -975,13 +1094,13 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
 * Update connections
 */
 void
-HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation)
+HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int m, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation)
 {
-	HnswNeighborArray *currentNeighbors = &hc->element->neighbors[lc];
-
+	HnswElement hce = HnswPtrAccess(base, hc->element);
+	HnswNeighborArray *currentNeighbors = HnswGetNeighbors(base, hce, lc);
 	HnswCandidate hc2;

-	hc2.element = element;
+	HnswPtrStore(base, hc2.element, element);
 	hc2.distance = hc->distance;

 	if (currentNeighbors->length < m)
@@ -1000,19 +1119,20 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = hc->element->value;
+			Datum		q = HnswGetValue(base, hce);

 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
+				HnswElement hc3Element = HnswPtrAccess(base, hc3->element);

-				if (DatumGetPointer(hc3->element->value) == NULL)
-					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
+				if (HnswPtrIsNull(base, hc3Element->value))
+					HnswLoadElement(hc3Element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
-					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
+					hc3->distance = GetCandidateDistance(base, hc3, q, procinfo, collation);

 				/* Prune element if being deleted */
-				if (list_length(hc3->element->heaptids) == 0)
+				if (hc3Element->heaptidsLength == 0)
 				{
 					pruned = &currentNeighbors->items[i];
 					break;
@@ -1029,7 +1149,7 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 				c = lappend(c, &currentNeighbors->items[i]);
 			c = lappend(c, &hc2);

-			SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
+			SelectNeighbors(base, c, m, lc, procinfo, collation, hce, &hc2, &pruned, true);

 			/* Should not happen */
 			if (pruned == NULL)
@@ -1039,7 +1159,7 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Find and replace the pruned element */
 		for (int i = 0; i < currentNeighbors->length; i++)
 		{
-			if (currentNeighbors->items[i].element == pruned->element)
+			if (HnswPtrEqual(base, currentNeighbors->items[i].element, pruned->element))
 			{
 				currentNeighbors->items[i] = hc2;

@@ -1057,43 +1177,65 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 * Remove elements being deleted or skipped
 */
 static List *
-RemoveElements(List *w, HnswElement skipElement)
+RemoveElements(char *base, List *w, HnswElement skipElement)
 {
 	ListCell   *lc2;
 	List	   *w2 = NIL;

+	/* Ensure does not access heaptidsLength during in-memory build */
+	pg_memory_barrier();
+
 	foreach(lc2, w)
 	{
 		HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
+		HnswElement hce = HnswPtrAccess(base, hc->element);

 		/* Skip self for vacuuming update */
-		if (skipElement != NULL && hc->element->blkno == skipElement->blkno && hc->element->offno == skipElement->offno)
+		if (skipElement != NULL && hce->blkno == skipElement->blkno && hce->offno == skipElement->offno)
 			continue;

-		if (list_length(hc->element->heaptids) != 0)
+		if (hce->heaptidsLength != 0)
 			w2 = lappend(w2, hc);
 	}

 	return w2;
 }

+#if PG_VERSION_NUM >= 130000
+/*
+ * Precompute hash
+ */
+static void
+PrecomputeHash(char *base, HnswElement element)
+{
+	HnswElementPtr ptr;
+
+	HnswPtrStore(base, ptr, element);
+
+	if (base == NULL)
+		element->hash = hash_pointer((uintptr_t) HnswPtrPointer(ptr));
+	else
+		element->hash = hash_offset(HnswPtrOffset(ptr));
+}
+#endif
+
 /*
 * Algorithm 1 from paper
 */
 void
-HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing)
+HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing)
 {
 	List	   *ep;
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = element->value;
+	Datum		q = HnswGetValue(base, element);
 	HnswElement skipElement = existing ? element : NULL;

 #if PG_VERSION_NUM >= 130000
 	/* Precompute hash */
 	if (index == NULL)
-		element->hash = hash_pointer((uintptr_t) element);
+		PrecomputeHash(base, element);
 #endif

 	/* No neighbors if no entry point */
@@ -1101,13 +1243,13 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		return;

 	/* Get entry point and level */
-	ep = list_make1(HnswEntryCandidate(entryPoint, q, index, procinfo, collation, true));
+	ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, procinfo, collation, true));
 	entryLevel = entryPoint->level;

 	/* 1st phase: greedy search to insert level */
 	for (int lc = entryLevel; lc >= level + 1; lc--)
 	{
-		w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
 		ep = w;
 	}

@@ -1125,12 +1267,12 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		List	   *neighbors;
 		List	   *lw;

-		w = HnswSearchLayer(q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
+		w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);

 		/* Elements being deleted or skipped can help with search */
 		/* but should be removed before selecting neighbors */
 		if (index != NULL)
-			lw = RemoveElements(w, skipElement);
+			lw = RemoveElements(base, w, skipElement);
 		else
 			lw = w;

@@ -1139,9 +1281,9 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		 * 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);
+		neighbors = SelectNeighbors(base, lw, lm, lc, procinfo, collation, element, NULL, NULL, false);

-		AddConnections(element, neighbors, lm, lc);
+		AddConnections(base, element, neighbors, lm, lc);

 		ep = w;
 	}
--- a/src/hnswvacuum.c
+++ b/src/hnswvacuum.c
@@ -199,21 +199,22 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	BufferAccessStrategy bas = vacuumstate->bas;
 	HnswNeighborTuple ntup = vacuumstate->ntup;
 	Size		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
+	char	   *base = NULL;

 	/* Skip if element is entry point */
 	if (entryPoint != NULL && element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
 		return;

 	/* Init fields */
-	HnswInitNeighbors(element, m);
-	element->heaptids = NIL;
+	HnswInitNeighbors(base, element, m, NULL);
+	element->heaptidsLength = 0;

-	/* Add element to graph, skipping itself */
-	HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, true);
+	/* Find neighbors for element, skipping itself */
+	HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, true);

 	/* Update neighbor tuple */
 	/* Do this before getting page to minimize locking */
-	HnswSetNeighborTuple(ntup, element, m);
+	HnswSetNeighborTuple(base, ntup, element, m);

 	/* Get neighbor page */
 	buf = ReadBufferExtended(index, MAIN_FORKNUM, element->neighborPage, RBM_NORMAL, bas);
@@ -230,7 +231,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
 	UnlockReleaseBuffer(buf);

 	/* Update neighbors */
-	HnswUpdateNeighborPages(index, procinfo, collation, element, m, true, false);
+	HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, true, false);
 }

 /*
@@ -301,7 +302,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
 			{
 				/* Reset neighbors from previous update */
 				if (highestPoint != NULL)
-					highestPoint->neighbors = NULL;
+					HnswPtrStore((char *) NULL, highestPoint->neighbors, (HnswNeighborArrayPtr *) NULL);

 				RepairGraphElement(vacuumstate, entryPoint, highestPoint);
 			}
--- a/src/ivfbuild.c
+++ b/src/ivfbuild.c
@@ -543,10 +543,10 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	pfree(list);
 }

+#ifdef IVFFLAT_KMEANS_DEBUG
 /*
 * Print k-means metrics
 */
-#ifdef IVFFLAT_KMEANS_DEBUG
 static void
 PrintKmeansMetrics(IvfflatBuildState * buildstate)
 {
--- a/src/ivfkmeans.c
+++ b/src/ivfkmeans.c
@@ -6,6 +6,10 @@
 #include "ivfflat.h"
 #include "miscadmin.h"

+#ifdef IVFFLAT_MEMORY
+#include "utils/memutils.h"
+#endif
+
 /*
 * Initialize with kmeans++
 *
@@ -151,6 +155,23 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 	}
 }

+#ifdef IVFFLAT_MEMORY
+/*
+ * Show memory usage
+ */
+static void
+ShowMemoryUsage(Size estimatedSize)
+{
+#if PG_VERSION_NUM >= 130000
+	elog(INFO, "total memory: %zu MB",
+		 MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
+#else
+	MemoryContextStats(CurrentMemoryContext);
+#endif
+	elog(INFO, "estimated memory: %zu MB", estimatedSize / (1024 * 1024));
+}
+#endif
+
 /*
 * Use Elkan for performance. This requires distance function to satisfy triangle inequality.
 *
@@ -231,6 +252,10 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 		vec->dim = dimensions;
 	}

+#ifdef IVFFLAT_MEMORY
+	ShowMemoryUsage(totalSize);
+#endif
+
 	/* Pick initial centers */
 	InitCenters(index, samples, centers, lowerBound);

--- a/test/t/012_hnsw_build_recall.pl
+++ b/test/t/012_hnsw_build_recall.pl
@@ -95,11 +95,10 @@ for my $i (0 .. $#operators)

 	$node->safe_psql("postgres", "DROP INDEX idx;");

-	# Build index in parallel
+	# Build index in parallel in memory
 	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);
@@ -109,6 +108,21 @@ for my $i (0 .. $#operators)
 	test_recall($min, $operator);

 	$node->safe_psql("postgres", "DROP INDEX idx;");
+
+	# Build index in parallel on disk
+	# Set parallel_workers on table to use workers with low maintenance_work_mem
+	($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+		ALTER TABLE tst SET (parallel_workers = 2);
+		SET client_min_messages = DEBUG;
+		SET maintenance_work_mem = '4MB';
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+		ALTER TABLE tst RESET (parallel_workers);
+	));
+	is($ret, 0, $stderr);
+	like($stderr, qr/using \d+ parallel workers/);
+	like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem/);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
 }

 done_testing();
Author	SHA1	Message	Date
Andrew Kane	13d3bae99b	Less space [skip ci]	2024-01-22 18:43:26 -08:00
Andrew Kane	3b9c4c55ee	Use consistent order [skip ci]	2024-01-22 18:39:52 -08:00
Andrew Kane	06e2052e40	Added InitVisited [skip ci]	2024-01-22 18:37:37 -08:00
Andrew Kane	891072df68	Added base [skip ci]	2024-01-22 18:02:24 -08:00
Andrew Kane	e9a96d341f	Removed prefix for static functions	2024-01-22 17:58:42 -08:00
Andrew Kane	a45ca414f4	Keep logic consistent with insert [skip ci]	2024-01-22 17:43:04 -08:00
Andrew Kane	b4f3cc3e13	Moved outside loop [skip ci]	2024-01-22 17:26:26 -08:00
Andrew Kane	c01cf8a315	Renamed ApplyChanges to UpdateGraph [skip ci]	2024-01-22 17:02:02 -08:00
Andrew Kane	3ecb9a3cb2	Renamed HnswInsertElement to HnswFindElementNeighbors [skip ci]	2024-01-22 16:59:08 -08:00
Andrew Kane	a069e18fe4	Improved function names [skip ci]	2024-01-22 16:48:50 -08:00
Andrew Kane	5174a23094	Updated comment [skip ci]	2024-01-22 16:45:42 -08:00
Andrew Kane	16d7de79f6	Improved function names [skip ci]	2024-01-22 16:43:50 -08:00
Andrew Kane	e54ec4d637	Improved code [skip ci]	2024-01-22 16:37:32 -08:00
Andrew Kane	cc641002d3	Updated comments [skip ci]	2024-01-22 10:59:46 -08:00
Heikki Linnakangas	2f9b1e2893	xAdd overview comment on how HNSW build works (#419 ) And rewrite some of the comments in InsertTuple(), to also give more of a high-level overview of the flow.	2024-01-22 10:50:12 -08:00
Andrew Kane	a3e4fbf6aa	Use shared lock for copying neighbors to local memory	2024-01-19 13:44:25 -08:00
Andrew Kane	09a4ec29a0	Added InsertTupleInMemory	2024-01-19 01:27:45 -08:00
Heikki Linnakangas	ca3b4cd029	Remove HnswSpool It was just used to pass heap/index relations to HnswParallelScanAndInsert. I think it was copied from nbtsort.c, which is more complicated. I don't think we need a struct like this. (That said, I actually think that we should have a state object that would hold fields like 'heap', 'index', 'procinfo', 'collation' etc. Passing that object around would simplify the signatures of many functions. But that's a different story).	2024-01-19 00:26:47 -08:00
Heikki Linnakangas	d96e486274	Remove unused 'scantuplesortstates' field	2024-01-19 00:26:47 -08:00
Heikki Linnakangas	88213186a5	Remove unused argument	2024-01-19 00:26:47 -08:00
Andrew Kane	7dd9534894	Use same locking as insert	2024-01-19 00:18:29 -08:00
Andrew Kane	d801a843f4	Removed HnswPtrSetNull to avoid setting relptr_off directly	2024-01-16 17:08:13 -08:00
Andrew Kane	1458c7bb2a	Improved code [skip ci]	2024-01-16 14:03:28 -08:00
Andrew Kane	cad48d9203	Improved locking	2024-01-16 13:34:55 -08:00
Heikki Linnakangas	719b4b7436	Use LWLocks instead of SpinLocks (#410 ) Spinlocks should be held only for a few instructions, for multiple reasons: - You have to be very careful not to elog() out while holding a spinlock, because there is no mechanism to release the spinlock on error. - Waiters can waste a lot of cycles spinning if the lock is contended. I you wait on a spinlock for too long, the PostgreSQL implementation will actually PANIC, see s_lock_stuck(). The flushLock is particularly problematic. It is held in exclusive mode, which means it holds a spinlock, over the call to FlushPages(). FlushPages() performs lots of I/O so it can take a very long time (>= minutes), and can also easily error out for various reasons. allocatorLock would perhaps be OK as a spinlocks, but even that feels a bit heavy, so I converted that to an LWLock, too. entryLock is usually held for a very short time, in shared mode, so that would be fine as a spinlock. However, in the rare case that the entry point is updated, it's held for a very long time. An LWLock used in shared mode is about as fast a spinlock, that path is pretty heavily optimized. I think we have some problems with the per-element spinlocks too. In HnswUpdateNeighborPagesInMemory(), it's held over a call to HnswUpdateConnection(), but HnswUpdateConnection() can error out at least in case of an out-of-memory error (it uses lappend(), which calls palloc()). It also calls the distance function, and I don't think they are guaranteed to be ereport-free either. However, I didn't address that in this PR, it needs a bit more thinking.	2024-01-16 13:25:03 -08:00
Andrew Kane	fa0acbf62d	Fixed CI	2024-01-15 19:55:46 -08:00
Andrew Kane	1612b84069	Fixed error on Windows [skip ci]	2024-01-15 19:33:16 -08:00
Andrew Kane	2f9371516d	Leave space for other objects in shared memory	2024-01-15 19:17:50 -08:00
Andrew Kane	9d3e4e74df	Added support for in-memory parallel index builds for HNSW	2024-01-15 15:07:31 -08:00
Andrew Kane	0ce497a1b1	Updated Homebrew note [skip ci]	2024-01-15 12:12:04 -08:00
Andrew Kane	c7d60346d8	Improved macro [skip ci]	2024-01-13 20:02:41 -08:00
Andrew Kane	597bfdc76b	Added HnswGetNeighbors macro	2024-01-13 20:00:34 -08:00
Andrew Kane	cbf3eb4fa5	Improved HNSW build and insert code	2024-01-13 10:07:42 -08:00
Andrew Kane	cacd389f6d	Improved pattern for duplicates	2024-01-12 14:30:13 -08:00
Andrew Kane	423cc2b06c	Homebrew now adds to postgresql@15 as well [skip ci]	2024-01-11 16:45:50 -08:00
Andrew Kane	85c4ef6a14	Updated Postgres versions in readme [skip ci]	2024-01-11 12:36:24 -08:00
Andrew Kane	c6160a783a	Homebrew now adds to postgresql@16 [skip ci]	2024-01-11 12:32:14 -08:00
Andrew Kane	1881b857f9	Simplified code	2024-01-09 18:53:31 -08:00
Andrew Kane	51bde5fb22	Updated readme [skip ci]	2024-01-09 14:38:25 -08:00
Andrew Kane	10e65ce349	Added note about maintenance_work_mem [skip ci]	2024-01-09 14:31:54 -08:00
Andrew Kane	61279f5a59	Updated readme [skip ci]	2024-01-09 14:26:55 -08:00
Andrew Kane	72b3889e26	Updated readme [skip ci]	2024-01-09 14:22:19 -08:00
Andrew Kane	bb21b2decf	Updated readme [skip ci]	2024-01-09 14:19:01 -08:00
Andrew Kane	8a65c0e831	Moved section [skip ci]	2024-01-09 13:33:03 -08:00
Andrew Kane	7d75d423e4	Added section on index build time [skip ci]	2024-01-09 13:27:27 -08:00
Andrew Kane	6cad1f5de0	Updated example [skip ci]	2024-01-09 13:04:47 -08:00
Andrew Kane	67eeade63c	Moved HNSW first in readme [skip ci]	2024-01-09 13:04:18 -08:00
Andrew Kane	108fb09d7b	Improved code [skip ci]	2024-01-08 17:54:49 -08:00
Andrew Kane	65d060ac86	Reverted FlushPages pattern for parallel builds	2024-01-08 10:45:31 -08:00
Andrew Kane	62ee33bb92	Improved locking code	2024-01-08 09:05:12 -08:00
Andrew Kane	520e274dde	Improved locking code	2024-01-07 22:34:41 -08:00
Andrew Kane	9e680884bd	Moved indtuples to HnswGraph	2024-01-07 22:23:49 -08:00
Andrew Kane	19a0e1b341	Moved graph to separate struct	2024-01-07 20:15:30 -08:00
Andrew Kane	c7fe1571ee	Improved code	2024-01-07 18:30:51 -08:00
Andrew Kane	cb4c770df2	Switched to slist for elements to reduce allocations and remove limit	2024-01-07 18:26:19 -08:00
Andrew Kane	85fdecd79b	Moved FlushPages before HnswEndParallel	2024-01-07 17:50:46 -08:00
Andrew Kane	6132428914	Improved number of parallel workers for HNSW index builds - closes #397	2024-01-05 19:46:08 -08:00
Andrew Kane	81d13bd40f	Improved code [skip ci]	2024-01-03 13:53:23 -05:00
Andrew Kane	8ee37b60a0	Improved memory estimate for HNSW index builds	2024-01-03 13:47:50 -05:00
Andrew Kane	9b73b3d1a6	Reduced memory and allocations for heap TIDs - closes #385	2024-01-03 13:41:34 -05:00
Andrew Kane	cae630784b	Improved BuildCallback [skip ci]	2023-12-30 20:55:29 -05:00
Andrew Kane	d87bcd2deb	Added comments [skip ci]	2023-12-30 18:29:01 -05:00
Andrew Kane	736576220a	Improved BuildCallback	2023-12-30 18:24:03 -05:00
Andrew Kane	a508b120c1	Added IVFFLAT_MEMORY flag to show memory usage [skip ci]	2023-12-24 09:27:09 -05:00
Andrew Kane	9a782d29f8	Use consistent style [skip ci]	2023-12-22 16:41:25 -05:00
Andrew Kane	1e422cd62b	Improved readability [skip ci]	2023-12-22 16:39:13 -05:00
Andrew Kane	569c69580a	Improved InsertTuple code - #384 Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>	2023-12-22 15:08:28 -05:00
Andrew Kane	59509c3a17	Added extra 5% to memory estimate	2023-12-22 14:04:05 -05:00
Andrew Kane	61738846af	Updated comment [skip ci]	2023-12-22 14:03:33 -05:00
Andrew Kane	e8c3bf0cef	Improved memory tracking for HNSW index builds - #384	2023-12-22 13:35:43 -05:00
Andrew Kane	50d1aed3d8	Improved memory usage logging [skip ci]	2023-12-22 13:09:11 -05:00
Andrew Kane	66e14d2434	Updated indentation [skip ci]	2023-12-22 12:59:50 -05:00
Andrew Kane	42cd4c6833	Fixed call to GenerationContextCreate for Postgres < 15	2023-12-22 12:49:07 -05:00
Andrew Kane	dcbe0b6f0d	Reduced memory usage for HNSW index builds - #384 Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>	2023-12-22 12:41:47 -05:00
Andrew Kane	f61d4087b5	Slightly improved memory estimation [skip ci]	2023-12-21 10:31:36 -05:00
Andrew Kane	57554e5b46	Added todo [skip ci]	2023-12-20 17:52:31 -05:00
Andrew Kane	6738fa0bd7	Added HNSW_MEMORY flag to show memory usage - #384 [skip ci]	2023-12-20 16:49:16 -05:00
Andrew Kane	9ab10aa674	Fixed CI	2023-12-20 16:29:13 -05:00
Andrew Kane	ec41dfa1d7	Mark meta buffer contents as dirty when not logging	2023-12-20 16:20:15 -05:00
Andrew Kane	43e0b3d9d4	Mark buffer contents as dirty when not logging	2023-12-20 16:16:25 -05:00
Andrew Kane	2bff7ccaa2	Improved memory estimation - #384 [skip ci]	2023-12-20 11:10:27 -05:00
Andrew Kane	e88a425c9b	Reduced WAL generation for HNSW index builds - thanks @hlinnaka	2023-12-19 20:37:32 -05:00