Add HNSW extension (#4227)

## Describe your changes Port HNSW implementation for ANN search top Postgres ## Issue ticket number and link https://www.pinecone.io/learn/hnsw ## Checklist before requesting a review - [ ] I have performed a self-review of my code. - [ ] If it is a core feature, I have added thorough tests. - [ ] Do we need to implement analytics? if so did you add the relevant metrics to the dashboard? - [ ] If this PR requires public announcement, mark it with /release-notes label and add several sentences in this section. ## Checklist before merging - [ ] Do not forget to reformat commit message to not include the above checklist
2026-01-05 20:42:54 +00:00 · 2023-06-04 11:41:38 +03:00
parent 4ba950a35a
commit 04542826be
12 changed files with 1139 additions and 0 deletions
--- a/Dockerfile.compute-node
+++ b/Dockerfile.compute-node
@@ -556,6 +556,10 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) \
    make -j $(getconf _NPROCESSORS_ONLN) \
        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
        -C pgxn/neon_utils \
+        -s install && \
+    make -j $(getconf _NPROCESSORS_ONLN) \
+        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
+        -C pgxn/hnsw \
        -s install

 #########################################################################################
--- a/8
+++ b/8
@@ -138,6 +138,11 @@ neon-pg-ext-%: postgres-%
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile install
+	+@echo "Compiling hnsw $*"
+	mkdir -p $(POSTGRES_INSTALL_DIR)/build/hnsw-$*
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		-C $(POSTGRES_INSTALL_DIR)/build/hnsw-$* \
+		-f $(ROOT_PROJECT_DIR)/pgxn/hnsw/Makefile install

 .PHONY: neon-pg-ext-clean-%
 neon-pg-ext-clean-%:
@@ -153,6 +158,9 @@ neon-pg-ext-clean-%:
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
 	-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 	-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile clean
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
+	-C $(POSTGRES_INSTALL_DIR)/build/hnsw-$* \
+	-f $(ROOT_PROJECT_DIR)/pgxn/hnsw/Makefile clean

 .PHONY: neon-pg-ext
 neon-pg-ext: \
--- a/pgxn/hnsw/Makefile
+++ b/pgxn/hnsw/Makefile
@@ -0,0 +1,26 @@
+EXTENSION = hnsw
+EXTVERSION = 0.1.0
+
+MODULE_big = hnsw
+DATA = $(wildcard *--*.sql)
+OBJS = hnsw.o hnswalg.o
+
+TESTS = $(wildcard test/sql/*.sql)
+REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
+REGRESS_OPTS = --inputdir=test --load-extension=hnsw
+
+# For auto-vectorization:
+# - GCC (needs -ftree-vectorize OR -O3) - https://gcc.gnu.org/projects/tree-ssa/vectorization.html
+PG_CFLAGS += -O3
+PG_CPPFLAGS +=  -msse4.1 -O3 -march=native -ftree-vectorize -ftree-vectorizer-verbose=0
+PG_LDFLAGS += -lstdc++
+
+all: $(EXTENSION)--$(EXTVERSION).sql
+
+PG_CONFIG ?= pg_config
+PGXS := $(shell $(PG_CONFIG) --pgxs)
+include $(PGXS)
+
+dist:
+	mkdir -p dist
+	git archive --format zip --prefix=$(EXTENSION)-$(EXTVERSION)/ --output dist/$(EXTENSION)-$(EXTVERSION).zip master
--- a/pgxn/hnsw/README.md
+++ b/pgxn/hnsw/README.md
@@ -0,0 +1,25 @@
+# Revisiting the Inverted Indices for Billion-Scale Approximate Nearest Neighbors
+
+This ANN extension of Postgres is based
+on [ivf-hnsw](https://github.com/dbaranchuk/ivf-hnsw.git) implementation of [HNSW](https://www.pinecone.io/learn/hnsw),
+the code for the current state-of-the-art billion-scale nearest neighbor search system presented in the paper:
+
+[Revisiting the Inverted Indices for Billion-Scale Approximate Nearest Neighbors](http://openaccess.thecvf.com/content_ECCV_2018/html/Dmitry_Baranchuk_Revisiting_the_Inverted_ECCV_2018_paper.html),
+<br>
+Dmitry Baranchuk, Artem Babenko, Yury Malkov
+
+# Postgres extension
+
+HNSW index is hold in memory (built on demand) and it's maxial size is limited
+by `maxelements` index parameter. Another required parameter is nubmer of dimensions (if it is not specified in column type).
+Optional parameter `ef` specifies number of neighbors which are considered during index construction and search (corresponds `efConstruction` and `efSearch` parameters
+described in the article).
+
+# Example of usage:
+
+```
+create extension hnsw;
+create table embeddings(id integer primary key, payload real[]);
+create index on embeddings using hnsw(payload) with (maxelements=1000000, dims=100, m=32);
+select id from embeddings order by payload <-> array[1.0, 2.0,...] limit 100;
+```
--- a/pgxn/hnsw/hnsw--0.1.0.sql
+++ b/pgxn/hnsw/hnsw--0.1.0.sql
@@ -0,0 +1,29 @@
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "CREATE EXTENSION hnsw" to load this file. \quit
+
+-- functions
+
+CREATE FUNCTION l2_distance(real[], real[]) RETURNS real
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- operators
+
+CREATE OPERATOR <-> (
+	LEFTARG = real[], RIGHTARG = real[], PROCEDURE = l2_distance,
+	COMMUTATOR = '<->'
+);
+
+-- access method
+
+CREATE FUNCTION hnsw_handler(internal) RETURNS index_am_handler
+	AS 'MODULE_PATHNAME' LANGUAGE C;
+
+CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnsw_handler;
+
+COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
+
+-- opclasses
+
+CREATE OPERATOR CLASS knn_ops
+	DEFAULT FOR TYPE real[] USING hnsw AS
+	OPERATOR 1 <-> (real[], real[]) FOR ORDER BY float_ops;
--- a/pgxn/hnsw/hnsw.c
+++ b/pgxn/hnsw/hnsw.c
@@ -0,0 +1,551 @@
+#include "postgres.h"
+
+#include "access/amapi.h"
+#include "access/generic_xlog.h"
+#include "access/relation.h"
+#include "access/reloptions.h"
+#include "access/tableam.h"
+#include "catalog/index.h"
+#include "commands/vacuum.h"
+#include "nodes/execnodes.h"
+#include "storage/bufmgr.h"
+#include "utils/guc.h"
+#include "utils/selfuncs.h"
+
+#include <math.h>
+#include <float.h>
+
+#include "hnsw.h"
+
+PG_MODULE_MAGIC;
+
+typedef struct {
+	int32 vl_len_;		/* varlena header (do not touch directly!) */
+	int dims;
+	int maxelements;
+	int efConstruction;
+	int efSearch;
+	int M;
+} HnswOptions;
+
+static relopt_kind hnsw_relopt_kind;
+
+typedef struct {
+	HierarchicalNSW* hnsw;
+	size_t curr;
+	size_t n_results;
+	ItemPointer results;
+} HnswScanOpaqueData;
+
+typedef HnswScanOpaqueData* HnswScanOpaque;
+
+typedef struct {
+	Oid relid;
+	uint32 status;
+	HierarchicalNSW* hnsw;
+} HnswHashEntry;
+
+
+#define SH_PREFIX			 hnsw_index
+#define SH_ELEMENT_TYPE		 HnswHashEntry
+#define SH_KEY_TYPE			 Oid
+#define SH_KEY				 relid
+#define SH_STORE_HASH
+#define SH_GET_HASH(tb, a)	 ((a)->relid)
+#define SH_HASH_KEY(tb, key) (key)
+#define SH_EQUAL(tb, a, b)	((a) == (b))
+#define SH_SCOPE			static inline
+#define SH_DEFINE
+#define SH_DECLARE
+#include "lib/simplehash.h"
+
+#define INDEX_HASH_SIZE     11
+
+#define DEFAULT_EF_SEARCH   64
+
+PGDLLEXPORT void _PG_init(void);
+
+static hnsw_index_hash *hnsw_indexes;
+
+/*
+ * Initialize index options and variables
+ */
+void
+_PG_init(void)
+{
+	hnsw_relopt_kind = add_reloption_kind();
+	add_int_reloption(hnsw_relopt_kind, "dims", "Number of dimensions",
+					  0, 0, INT_MAX, AccessExclusiveLock);
+	add_int_reloption(hnsw_relopt_kind, "maxelements", "Maximal number of elements",
+					  0, 0, INT_MAX, AccessExclusiveLock);
+	add_int_reloption(hnsw_relopt_kind, "m", "Number of neighbors of each vertex",
+					  100, 0, INT_MAX, AccessExclusiveLock);
+	add_int_reloption(hnsw_relopt_kind, "efconstruction", "Number of inspected neighbors during index construction",
+					  16, 1, INT_MAX, AccessExclusiveLock);
+	add_int_reloption(hnsw_relopt_kind, "efsearch", "Number of inspected neighbors during index search",
+					  64, 1, INT_MAX, AccessExclusiveLock);
+	hnsw_indexes = hnsw_index_create(TopMemoryContext, INDEX_HASH_SIZE, NULL);
+}
+
+
+static void
+hnsw_build_callback(Relation index, ItemPointer tid, Datum *values,
+					bool *isnull, bool tupleIsAlive, void *state)
+{
+	HierarchicalNSW* hnsw = (HierarchicalNSW*) state;
+	ArrayType* array;
+	int n_items;
+	label_t label = 0;
+
+	/* Skip nulls */
+	if (isnull[0])
+		return;
+
+	array = DatumGetArrayTypeP(values[0]);
+	n_items = ArrayGetNItems(ARR_NDIM(array), ARR_DIMS(array));
+	if (n_items != hnsw_dimensions(hnsw))
+	{
+		elog(ERROR, "Wrong number of dimensions: %d instead of %d expected",
+			 n_items, hnsw_dimensions(hnsw));
+	}
+
+	memcpy(&label, tid, sizeof(*tid));
+	hnsw_add_point(hnsw, (coord_t*)ARR_DATA_PTR(array), label);
+}
+
+static void
+hnsw_populate(HierarchicalNSW* hnsw, Relation indexRel, Relation heapRel)
+{
+	IndexInfo* indexInfo = BuildIndexInfo(indexRel);
+	Assert(indexInfo->ii_NumIndexAttrs == 1);
+	table_index_build_scan(heapRel, indexRel, indexInfo,
+						   true, true, hnsw_build_callback, (void *) hnsw, NULL);
+}
+
+static HierarchicalNSW*
+hnsw_get_index(Relation indexRel, Relation heapRel)
+{
+	HierarchicalNSW* hnsw;
+	Oid indexoid = RelationGetRelid(indexRel);
+	HnswHashEntry* entry = hnsw_index_lookup(hnsw_indexes, indexoid);
+	if (entry == NULL)
+	{
+		size_t dims, maxelements;
+		size_t M;
+		size_t maxM;
+		size_t size_links_level0;
+		size_t size_data_per_element;
+		size_t data_size;
+		dsm_handle handle = indexoid << 1; /* make it even */
+		void* impl_private = NULL;
+		void* mapped_address = NULL;
+		Size  mapped_size = 0;
+		Size  shmem_size;
+		bool exists = true;
+		bool found;
+		HnswOptions *opts = (HnswOptions *) indexRel->rd_options;
+		if (opts == NULL || opts->maxelements == 0 || opts->dims == 0) {
+			elog(ERROR, "HNSW index requires 'maxelements' and 'dims' to be specified");
+		}
+		dims = opts->dims;
+		maxelements = opts->maxelements;
+		M = opts->M;
+		maxM = M * 2;
+		data_size = dims * sizeof(coord_t);
+		size_links_level0 = (maxM + 1) * sizeof(idx_t);
+		size_data_per_element = size_links_level0 + data_size + sizeof(label_t);
+		shmem_size =  hnsw_sizeof() + maxelements * size_data_per_element;
+
+		/* first try to attach to existed index */
+		if (!dsm_impl_op(DSM_OP_ATTACH, handle, 0, &impl_private,
+						 &mapped_address, &mapped_size, DEBUG1))
+		{
+			/* index doesn't exists: try to create it */
+			if (!dsm_impl_op(DSM_OP_CREATE, handle, shmem_size, &impl_private,
+							 &mapped_address, &mapped_size, DEBUG1))
+			{
+				/* We can do it under shared lock, so some other backend may
+				 * try to initialize index. If create is failed because index already
+				 * created by somebody else, then try to attach to it once again
+				 */
+				if (!dsm_impl_op(DSM_OP_ATTACH, handle, 0, &impl_private,
+								 &mapped_address, &mapped_size, ERROR))
+				{
+					return NULL;
+				}
+			}
+			else
+			{
+				exists = false;
+			}
+		}
+		Assert(mapped_size == shmem_size);
+		hnsw = (HierarchicalNSW*)mapped_address;
+
+		if (!exists)
+		{
+			hnsw_init(hnsw, dims, maxelements, M, maxM, opts->efConstruction);
+			hnsw_populate(hnsw, indexRel, heapRel);
+		}
+		entry = hnsw_index_insert(hnsw_indexes, indexoid, &found);
+		Assert(!found);
+		entry->hnsw = hnsw;
+	}
+	else
+	{
+		hnsw = entry->hnsw;
+	}
+	return hnsw;
+}
+
+/*
+ * Start or restart an index scan
+ */
+static IndexScanDesc
+hnsw_beginscan(Relation index, int nkeys, int norderbys)
+{
+	IndexScanDesc scan = RelationGetIndexScan(index, nkeys, norderbys);
+	HnswScanOpaque so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
+	Relation heap = relation_open(index->rd_index->indrelid, NoLock);
+	so->hnsw = hnsw_get_index(index, heap);
+	relation_close(heap, NoLock);
+	so->curr = 0;
+	so->n_results = 0;
+	so->results = NULL;
+	scan->opaque = so;
+	return scan;
+}
+
+/*
+ * Start or restart an index scan
+ */
+static void
+hnsw_rescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int norderbys)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	if (so->results)
+	{
+		pfree(so->results);
+		so->results = NULL;
+	}
+	so->curr = 0;
+	if (orderbys && scan->numberOfOrderBys > 0)
+		memmove(scan->orderByData, orderbys, scan->numberOfOrderBys * sizeof(ScanKeyData));
+}
+
+/*
+ * Fetch the next tuple in the given scan
+ */
+static bool
+hnsw_gettuple(IndexScanDesc scan, ScanDirection dir)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+
+	/*
+	 * Index can be used to scan backward, but Postgres doesn't support
+	 * backward scan on operators
+	 */
+	Assert(ScanDirectionIsForward(dir));
+
+	if (so->curr == 0)
+	{
+		Datum		value;
+		ArrayType*	array;
+		int         n_items;
+		size_t      n_results;
+		label_t*    results;
+		HnswOptions *opts = (HnswOptions *) scan->indexRelation->rd_options;
+		size_t      efSearch = opts ? opts->efSearch : DEFAULT_EF_SEARCH;
+
+		/* Safety check */
+		if (scan->orderByData == NULL)
+			elog(ERROR, "cannot scan HNSW index without order");
+
+		/* No items will match if null */
+		if (scan->orderByData->sk_flags & SK_ISNULL)
+			return false;
+
+		value = scan->orderByData->sk_argument;
+		array = DatumGetArrayTypeP(value);
+		n_items = ArrayGetNItems(ARR_NDIM(array), ARR_DIMS(array));
+		if (n_items != hnsw_dimensions(so->hnsw))
+		{
+			elog(ERROR, "Wrong number of dimensions: %d instead of %d expected",
+				 n_items, hnsw_dimensions(so->hnsw));
+		}
+
+		if (!hnsw_search(so->hnsw, (coord_t*)ARR_DATA_PTR(array), efSearch, &n_results, &results))
+			elog(ERROR, "HNSW index search failed");
+		so->results = (ItemPointer)palloc(n_results*sizeof(ItemPointerData));
+		so->n_results = n_results;
+		for (size_t i = 0; i < n_results; i++)
+		{
+			memcpy(&so->results[i], &results[i], sizeof(so->results[i]));
+		}
+		free(results);
+	}
+	if (so->curr >= so->n_results)
+	{
+		return false;
+	}
+	else
+	{
+		scan->xs_heaptid = so->results[so->curr++];
+		scan->xs_recheckorderby = false;
+		return true;
+	}
+}
+
+/*
+ * End a scan and release resources
+ */
+static void
+hnsw_endscan(IndexScanDesc scan)
+{
+	HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
+	if (so->results)
+		pfree(so->results);
+	pfree(so);
+	scan->opaque = NULL;
+}
+
+
+/*
+ * Estimate the cost of an index scan
+ */
+static void
+hnsw_costestimate(PlannerInfo *root, IndexPath *path, double loop_count,
+				 Cost *indexStartupCost, Cost *indexTotalCost,
+				 Selectivity *indexSelectivity, double *indexCorrelation
+				 ,double *indexPages
+)
+{
+	GenericCosts costs;
+
+	/* Never use index without order */
+	if (path->indexorderbys == NULL)
+	{
+		*indexStartupCost = DBL_MAX;
+		*indexTotalCost = DBL_MAX;
+		*indexSelectivity = 0;
+		*indexCorrelation = 0;
+		*indexPages = 0;
+		return;
+	}
+
+	MemSet(&costs, 0, sizeof(costs));
+
+	genericcostestimate(root, path, loop_count, &costs);
+
+	/* Startup cost and total cost are same */
+	*indexStartupCost = costs.indexTotalCost;
+	*indexTotalCost = costs.indexTotalCost;
+	*indexSelectivity = costs.indexSelectivity;
+	*indexCorrelation = costs.indexCorrelation;
+	*indexPages = costs.numIndexPages;
+}
+
+/*
+ * Parse and validate the reloptions
+ */
+static bytea *
+hnsw_options(Datum reloptions, bool validate)
+{
+	static const relopt_parse_elt tab[] = {
+		{"dims", RELOPT_TYPE_INT, offsetof(HnswOptions, dims)},
+		{"maxelements", RELOPT_TYPE_INT, offsetof(HnswOptions, maxelements)},
+		{"efconstruction", RELOPT_TYPE_INT, offsetof(HnswOptions, efConstruction)},
+		{"efsearch", RELOPT_TYPE_INT, offsetof(HnswOptions, efSearch)},
+		{"m", RELOPT_TYPE_INT, offsetof(HnswOptions, M)}
+	};
+
+	return (bytea *) build_reloptions(reloptions, validate,
+									  hnsw_relopt_kind,
+									  sizeof(HnswOptions),
+									  tab, lengthof(tab));
+}
+
+/*
+ * Validate catalog entries for the specified operator class
+ */
+static bool
+hnsw_validate(Oid opclassoid)
+{
+	return true;
+}
+
+/*
+ * Build the index for a logged table
+ */
+static IndexBuildResult *
+hnsw_build(Relation heap, Relation index, IndexInfo *indexInfo)
+{
+	HierarchicalNSW* hnsw = hnsw_get_index(index, heap);
+	IndexBuildResult* result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
+	result->heap_tuples = result->index_tuples = hnsw_count(hnsw);
+
+	return result;
+}
+
+/*
+ * Insert a tuple into the index
+ */
+static bool
+hnsw_insert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
+			  Relation heap, IndexUniqueCheck checkUnique,
+			  bool indexUnchanged,
+			  IndexInfo *indexInfo)
+{
+	HierarchicalNSW* hnsw = hnsw_get_index(index, heap);
+	Datum value;
+	ArrayType* array;
+	int n_items;
+	label_t label = 0;
+
+	/* Skip nulls */
+	if (isnull[0])
+		return false;
+
+	/* Detoast value */
+	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+	array = DatumGetArrayTypeP(value);
+	n_items = ArrayGetNItems(ARR_NDIM(array), ARR_DIMS(array));
+	if (n_items != hnsw_dimensions(hnsw))
+	{
+		elog(ERROR, "Wrong number of dimensions: %d instead of %d expected",
+			 n_items, hnsw_dimensions(hnsw));
+	}
+	memcpy(&label, heap_tid, sizeof(*heap_tid));
+	if (!hnsw_add_point(hnsw, (coord_t*)ARR_DATA_PTR(array), label))
+		elog(ERROR, "HNSW index insert failed");
+	return true;
+}
+
+/*
+ * Build the index for an unlogged table
+ */
+static void
+hnsw_buildempty(Relation index)
+{
+	/* index will be constructed on dema nd when accessed */
+}
+
+/*
+ * Clean up after a VACUUM operation
+ */
+static IndexBulkDeleteResult *
+hnsw_vacuumcleanup(IndexVacuumInfo *info, IndexBulkDeleteResult *stats)
+{
+	Relation	rel = info->index;
+
+	if (stats == NULL)
+		return NULL;
+
+	stats->num_pages = RelationGetNumberOfBlocks(rel);
+
+	return stats;
+}
+
+/*
+ * Bulk delete tuples from the index
+ */
+static IndexBulkDeleteResult *
+hnsw_bulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
+				IndexBulkDeleteCallback callback, void *callback_state)
+{
+	if (stats == NULL)
+		stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
+	return stats;
+}
+
+/*
+ * Define index handler
+ *
+ * See https://www.postgresql.org/docs/current/index-api.html
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(hnsw_handler);
+Datum
+hnsw_handler(PG_FUNCTION_ARGS)
+{
+	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
+
+	amroutine->amstrategies = 0;
+	amroutine->amsupport = 0;
+	amroutine->amoptsprocnum = 0;
+	amroutine->amcanorder = false;
+	amroutine->amcanorderbyop = true;
+	amroutine->amcanbackward = false;	/* can change direction mid-scan */
+	amroutine->amcanunique = false;
+	amroutine->amcanmulticol = false;
+	amroutine->amoptionalkey = true;
+	amroutine->amsearcharray = false;
+	amroutine->amsearchnulls = false;
+	amroutine->amstorage = false;
+	amroutine->amclusterable = false;
+	amroutine->ampredlocks = false;
+	amroutine->amcanparallel = false;
+	amroutine->amcaninclude = false;
+	amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
+	amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
+	amroutine->amkeytype = InvalidOid;
+
+	/* Interface functions */
+	amroutine->ambuild = hnsw_build;
+	amroutine->ambuildempty = hnsw_buildempty;
+	amroutine->aminsert = hnsw_insert;
+	amroutine->ambulkdelete = hnsw_bulkdelete;
+	amroutine->amvacuumcleanup = hnsw_vacuumcleanup;
+	amroutine->amcanreturn = NULL;	/* tuple not included in heapsort */
+	amroutine->amcostestimate = hnsw_costestimate;
+	amroutine->amoptions = hnsw_options;
+	amroutine->amproperty = NULL;	/* TODO AMPROP_DISTANCE_ORDERABLE */
+	amroutine->ambuildphasename = NULL;
+	amroutine->amvalidate = hnsw_validate;
+	amroutine->amadjustmembers = NULL;
+	amroutine->ambeginscan = hnsw_beginscan;
+	amroutine->amrescan = hnsw_rescan;
+	amroutine->amgettuple = hnsw_gettuple;
+	amroutine->amgetbitmap = NULL;
+	amroutine->amendscan = hnsw_endscan;
+	amroutine->ammarkpos = NULL;
+	amroutine->amrestrpos = NULL;
+
+	/* Interface functions to support parallel index scans */
+	amroutine->amestimateparallelscan = NULL;
+	amroutine->aminitparallelscan = NULL;
+	amroutine->amparallelrescan = NULL;
+
+	PG_RETURN_POINTER(amroutine);
+}
+
+/*
+ * Get the L2 distance between vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(l2_distance);
+Datum
+l2_distance(PG_FUNCTION_ARGS)
+{
+	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
+	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
+	int         a_dim = ArrayGetNItems(ARR_NDIM(a), ARR_DIMS(a));
+	int         b_dim = ArrayGetNItems(ARR_NDIM(b), ARR_DIMS(b));
+	dist_t 		distance = 0.0;
+	dist_t		diff;
+	coord_t	   *ax = (coord_t*)ARR_DATA_PTR(a);
+	coord_t	   *bx = (coord_t*)ARR_DATA_PTR(b);
+
+	if (a_dim != b_dim)
+	{
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("different array dimensions %d and %d", a_dim, b_dim)));
+	}
+
+	for (int i = 0; i < a_dim; i++)
+	{
+		diff = ax[i] - bx[i];
+		distance += diff * diff;
+	}
+
+	PG_RETURN_FLOAT4((dist_t)sqrt(distance));
+}
--- a/pgxn/hnsw/hnsw.control
+++ b/pgxn/hnsw/hnsw.control
@@ -0,0 +1,5 @@
+comment = 'hNsw index'
+default_version = '0.1.0'
+module_pathname = '$libdir/hnsw'
+relocatable = true
+trusted = true
--- a/pgxn/hnsw/hnsw.h
+++ b/pgxn/hnsw/hnsw.h
@@ -0,0 +1,15 @@
+#pragma once
+
+typedef float    coord_t;
+typedef float    dist_t;
+typedef uint32_t idx_t;
+typedef uint64_t label_t;
+
+typedef struct HierarchicalNSW HierarchicalNSW;
+
+bool hnsw_search(HierarchicalNSW* hnsw, const coord_t *point, size_t efSearch, size_t* n_results, label_t** results);
+bool hnsw_add_point(HierarchicalNSW* hnsw, const coord_t *point, label_t label);
+void hnsw_init(HierarchicalNSW* hnsw, size_t dim, size_t maxelements, size_t M, size_t maxM, size_t efConstruction);
+int  hnsw_dimensions(HierarchicalNSW* hnsw);
+size_t hnsw_count(HierarchicalNSW* hnsw);
+size_t hnsw_sizeof(void);
--- a/pgxn/hnsw/hnswalg.cpp
+++ b/pgxn/hnsw/hnswalg.cpp
@@ -0,0 +1,368 @@
+#include "hnswalg.h"
+
+
+#if defined(__x86_64__)
+
+#include <x86intrin.h>
+#define USE_AVX
+#if defined(__GNUC__)
+#define PORTABLE_ALIGN32 __attribute__((aligned(32)))
+#else
+#define PORTABLE_ALIGN32 __declspec(align(32))
+#endif
+
+#define PREFETCH(addr,hint) _mm_prefetch(addr, hint)
+
+#else
+
+#define PREFETCH(addr,hint)
+
+#endif
+
+HierarchicalNSW::HierarchicalNSW(size_t dim_, size_t maxelements_, size_t M_, size_t maxM_, size_t efConstruction_)
+{
+    dim = dim_;
+    data_size = dim * sizeof(coord_t);
+
+    efConstruction = efConstruction_;
+
+    maxelements = maxelements_;
+    M = M_;
+    maxM = maxM_;
+    size_links_level0 = (maxM + 1) * sizeof(idx_t);
+    size_data_per_element = size_links_level0 + data_size  + sizeof(label_t);
+    offset_data = size_links_level0;
+	offset_label = offset_data + data_size;
+
+    enterpoint_node = 0;
+    cur_element_count = 0;
+	dist_calc = 0;
+}
+
+std::priority_queue<std::pair<dist_t, idx_t>> HierarchicalNSW::searchBaseLayer(const coord_t *point, size_t ef)
+{
+	std::vector<uint32_t> visited;
+	visited.resize((cur_element_count + 31) >> 5);
+
+    std::priority_queue<std::pair<dist_t, idx_t >> topResults;
+    std::priority_queue<std::pair<dist_t, idx_t >> candidateSet;
+
+    dist_t dist = fstdistfunc(point, getDataByInternalId(enterpoint_node));
+
+    topResults.emplace(dist, enterpoint_node);
+    candidateSet.emplace(-dist, enterpoint_node);
+    visited[enterpoint_node >> 5] = 1 << (enterpoint_node & 31);
+    dist_t lowerBound = dist;
+
+    while (!candidateSet.empty())
+    {
+        std::pair<dist_t, idx_t> curr_el_pair = candidateSet.top();
+        if (-curr_el_pair.first > lowerBound)
+            break;
+
+        candidateSet.pop();
+        idx_t curNodeNum = curr_el_pair.second;
+
+        idx_t* data = get_linklist0(curNodeNum);
+        size_t size = *data++;
+
+        PREFETCH(getDataByInternalId(*data), _MM_HINT_T0);
+
+        for (size_t j = 0; j < size; ++j) {
+            size_t tnum = *(data + j);
+
+            PREFETCH(getDataByInternalId(*(data + j + 1)), _MM_HINT_T0);
+
+            if (!(visited[tnum >> 5] & (1 << (tnum & 31)))) {
+				visited[tnum >> 5] |= 1 << (tnum & 31);
+
+                dist = fstdistfunc(point, getDataByInternalId(tnum));
+
+                if (topResults.top().first > dist || topResults.size() < ef) {
+                    candidateSet.emplace(-dist, tnum);
+
+                    PREFETCH(get_linklist0(candidateSet.top().second), _MM_HINT_T0);
+                    topResults.emplace(dist, tnum);
+
+                    if (topResults.size() > ef)
+                        topResults.pop();
+
+                    lowerBound = topResults.top().first;
+                }
+            }
+        }
+    }
+    return topResults;
+}
+
+
+void HierarchicalNSW::getNeighborsByHeuristic(std::priority_queue<std::pair<dist_t, idx_t>> &topResults, size_t NN)
+{
+    if (topResults.size() < NN)
+        return;
+
+    std::priority_queue<std::pair<dist_t, idx_t>> resultSet;
+    std::vector<std::pair<dist_t, idx_t>> returnlist;
+
+    while (topResults.size() > 0) {
+        resultSet.emplace(-topResults.top().first, topResults.top().second);
+        topResults.pop();
+    }
+
+    while (resultSet.size()) {
+        if (returnlist.size() >= NN)
+            break;
+        std::pair<dist_t, idx_t> curen = resultSet.top();
+        dist_t dist_to_query = -curen.first;
+        resultSet.pop();
+        bool good = true;
+        for (std::pair<dist_t, idx_t> curen2 : returnlist) {
+            dist_t curdist = fstdistfunc(getDataByInternalId(curen2.second),
+                                         getDataByInternalId(curen.second));
+            if (curdist < dist_to_query) {
+                good = false;
+                break;
+            }
+        }
+        if (good) returnlist.push_back(curen);
+    }
+    for (std::pair<dist_t, idx_t> elem : returnlist)
+        topResults.emplace(-elem.first, elem.second);
+}
+
+void HierarchicalNSW::mutuallyConnectNewElement(const coord_t *point, idx_t cur_c,
+                               std::priority_queue<std::pair<dist_t, idx_t>> topResults)
+{
+    getNeighborsByHeuristic(topResults, M);
+
+    std::vector<idx_t> res;
+    res.reserve(M);
+    while (topResults.size() > 0) {
+        res.push_back(topResults.top().second);
+        topResults.pop();
+    }
+    {
+        idx_t* data = get_linklist0(cur_c);
+        if (*data)
+            throw std::runtime_error("Should be blank");
+
+        *data++ = res.size();
+
+        for (size_t idx = 0; idx < res.size(); idx++) {
+            if (data[idx])
+                throw std::runtime_error("Should be blank");
+            data[idx] = res[idx];
+        }
+    }
+    for (size_t idx = 0; idx < res.size(); idx++) {
+        if (res[idx] == cur_c)
+            throw std::runtime_error("Connection to the same element");
+
+        size_t resMmax = maxM;
+        idx_t *ll_other = get_linklist0(res[idx]);
+        idx_t sz_link_list_other = *ll_other;
+
+        if (sz_link_list_other > resMmax || sz_link_list_other < 0)
+            throw std::runtime_error("Bad sz_link_list_other");
+
+        if (sz_link_list_other < resMmax) {
+            idx_t *data = ll_other + 1;
+            data[sz_link_list_other] = cur_c;
+            *ll_other = sz_link_list_other + 1;
+        } else {
+            // finding the "weakest" element to replace it with the new one
+            idx_t *data = ll_other + 1;
+            dist_t d_max = fstdistfunc(getDataByInternalId(cur_c), getDataByInternalId(res[idx]));
+            // Heuristic:
+            std::priority_queue<std::pair<dist_t, idx_t>> candidates;
+            candidates.emplace(d_max, cur_c);
+
+            for (size_t j = 0; j < sz_link_list_other; j++)
+                candidates.emplace(fstdistfunc(getDataByInternalId(data[j]), getDataByInternalId(res[idx])), data[j]);
+
+            getNeighborsByHeuristic(candidates, resMmax);
+
+            size_t indx = 0;
+            while (!candidates.empty()) {
+                data[indx] = candidates.top().second;
+                candidates.pop();
+                indx++;
+            }
+            *ll_other = indx;
+        }
+    }
+}
+
+void HierarchicalNSW::addPoint(const coord_t *point, label_t label)
+{
+    if (cur_element_count >= maxelements) {
+        throw std::runtime_error("The number of elements exceeds the specified limit");
+    }
+    idx_t cur_c = cur_element_count++;
+    memset((char *) get_linklist0(cur_c), 0, size_data_per_element);
+    memcpy(getDataByInternalId(cur_c), point, data_size);
+    memcpy(getExternalLabel(cur_c), &label, sizeof label);
+
+    // Do nothing for the first element
+    if (cur_c != 0) {
+        std::priority_queue <std::pair<dist_t, idx_t>> topResults = searchBaseLayer(point, efConstruction);
+        mutuallyConnectNewElement(point, cur_c, topResults);
+    }
+};
+
+std::priority_queue<std::pair<dist_t, label_t>> HierarchicalNSW::searchKnn(const coord_t *query, size_t k)
+{
+	std::priority_queue<std::pair<dist_t, label_t>> topResults;
+	auto topCandidates = searchBaseLayer(query, k);
+    while (topCandidates.size() > k) {
+        topCandidates.pop();
+	}
+	while (!topCandidates.empty()) {
+		std::pair<dist_t, idx_t> rez = topCandidates.top();
+		label_t label;
+		memcpy(&label, getExternalLabel(rez.second), sizeof(label));
+		topResults.push(std::pair<dist_t, label_t>(rez.first, label));
+		topCandidates.pop();
+	}
+
+    return topResults;
+};
+
+dist_t HierarchicalNSW::fstdistfunc(const coord_t *x, const coord_t *y)
+{
+#if defined(__x86_64__)
+    float PORTABLE_ALIGN32 TmpRes[8];
+    size_t qty16 = dim >> 4;
+	const float *pEnd1 = x + (qty16 << 4);
+#ifdef USE_AVX
+	__m256 diff, v1, v2;
+	__m256 sum = _mm256_set1_ps(0);
+
+	while (x < pEnd1) {
+		v1 = _mm256_loadu_ps(x);
+		x += 8;
+		v2 = _mm256_loadu_ps(y);
+		y += 8;
+		diff = _mm256_sub_ps(v1, v2);
+		sum = _mm256_add_ps(sum, _mm256_mul_ps(diff, diff));
+
+		v1 = _mm256_loadu_ps(x);
+		x += 8;
+		v2 = _mm256_loadu_ps(y);
+		y += 8;
+		diff = _mm256_sub_ps(v1, v2);
+		sum = _mm256_add_ps(sum, _mm256_mul_ps(diff, diff));
+	}
+
+	_mm256_store_ps(TmpRes, sum);
+	float res = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3] + TmpRes[4] + TmpRes[5] + TmpRes[6] + TmpRes[7];
+
+	return (res);
+#else
+    __m128 diff, v1, v2;
+    __m128 sum = _mm_set1_ps(0);
+
+    while (x < pEnd1) {
+        v1 = _mm_loadu_ps(x);
+        x += 4;
+        v2 = _mm_loadu_ps(y);
+        y += 4;
+        diff = _mm_sub_ps(v1, v2);
+        sum = _mm_add_ps(sum, _mm_mul_ps(diff, diff));
+
+        v1 = _mm_loadu_ps(x);
+        x += 4;
+        v2 = _mm_loadu_ps(y);
+        y += 4;
+        diff = _mm_sub_ps(v1, v2);
+        sum = _mm_add_ps(sum, _mm_mul_ps(diff, diff));
+
+        v1 = _mm_loadu_ps(x);
+        x += 4;
+        v2 = _mm_loadu_ps(y);
+        y += 4;
+        diff = _mm_sub_ps(v1, v2);
+        sum = _mm_add_ps(sum, _mm_mul_ps(diff, diff));
+
+        v1 = _mm_loadu_ps(x);
+        x += 4;
+        v2 = _mm_loadu_ps(y);
+        y += 4;
+        diff = _mm_sub_ps(v1, v2);
+        sum = _mm_add_ps(sum, _mm_mul_ps(diff, diff));
+    }
+    _mm_store_ps(TmpRes, sum);
+    float res = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3];
+
+    return (res);
+#endif
+#else // portable implementation
+	dist_t 	distance = 0.0;
+	size_t  n = dim;
+
+	dist_calc++;
+
+	for (size_t i = 0; i < n; i++)
+	{
+		dist_t diff = x[i] - y[i];
+		distance += diff * diff;
+	}
+	return distance;
+#endif
+}
+
+bool hnsw_search(HierarchicalNSW* hnsw, const coord_t *point, size_t efSearch, size_t* n_results, label_t** results)
+{
+	try
+	{
+		auto result = hnsw->searchKnn(point, efSearch);
+		size_t nResults = result.size();
+		*results = (label_t*)malloc(nResults*sizeof(label_t));
+		for (size_t i = nResults; i-- != 0;)
+		{
+			(*results)[i] = result.top().second;
+			result.pop();
+		}
+		*n_results = nResults;
+		return true;
+	}
+	catch (std::exception& x)
+	{
+		return false;
+	}
+}
+
+bool hnsw_add_point(HierarchicalNSW* hnsw, const coord_t *point, label_t label)
+{
+	try
+	{
+		hnsw->addPoint(point, label);
+		return true;
+	}
+	catch (std::exception& x)
+	{
+		fprintf(stderr, "Catch %s\n", x.what());
+		return false;
+	}
+}
+
+void hnsw_init(HierarchicalNSW* hnsw, size_t dims, size_t maxelements, size_t M, size_t maxM, size_t efConstruction)
+{
+	new ((void*)hnsw) HierarchicalNSW(dims, maxelements, M, maxM, efConstruction);
+}
+
+
+int hnsw_dimensions(HierarchicalNSW* hnsw)
+{
+	return (int)hnsw->dim;
+}
+
+size_t hnsw_count(HierarchicalNSW* hnsw)
+{
+	return hnsw->cur_element_count;
+}
+
+size_t hnsw_sizeof(void)
+{
+	return sizeof(HierarchicalNSW);
+}
--- a/pgxn/hnsw/hnswalg.h
+++ b/pgxn/hnsw/hnswalg.h
@@ -0,0 +1,67 @@
+#pragma once
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdint.h>
+#include <unordered_map>
+#include <unordered_set>
+#include <map>
+#include <cmath>
+#include <queue>
+#include <stdexcept>
+
+extern "C" {
+#include "hnsw.h"
+}
+
+struct HierarchicalNSW
+{
+	size_t maxelements;
+	size_t cur_element_count;
+
+	idx_t  enterpoint_node;
+
+	size_t dist_calc;
+
+	size_t dim;
+	size_t data_size;
+	size_t offset_data;
+	size_t offset_label;
+	size_t size_data_per_element;
+	size_t M;
+	size_t maxM;
+	size_t size_links_level0;
+	size_t efConstruction;
+
+	char   data_level0_memory[0]; // varying size
+
+  public:
+	HierarchicalNSW(size_t dim, size_t maxelements, size_t M, size_t maxM, size_t efConstruction);
+	~HierarchicalNSW();
+
+
+	inline coord_t *getDataByInternalId(idx_t internal_id) const {
+		return (coord_t *)&data_level0_memory[internal_id * size_data_per_element + offset_data];
+	}
+
+	inline idx_t *get_linklist0(idx_t internal_id) const {
+		return (idx_t*)&data_level0_memory[internal_id * size_data_per_element];
+	}
+
+	inline label_t *getExternalLabel(idx_t internal_id) const {
+		return (label_t *)&data_level0_memory[internal_id * size_data_per_element + offset_label];
+	}
+
+	std::priority_queue<std::pair<dist_t, idx_t>> searchBaseLayer(const coord_t *x, size_t ef);
+
+	void getNeighborsByHeuristic(std::priority_queue<std::pair<dist_t, idx_t>> &topResults, size_t NN);
+
+	void mutuallyConnectNewElement(const coord_t *x, idx_t id, std::priority_queue<std::pair<dist_t, idx_t>> topResults);
+
+	void addPoint(const coord_t *point, label_t label);
+
+	std::priority_queue<std::pair<dist_t, label_t>> searchKnn(const coord_t *query_data, size_t k);
+
+	dist_t fstdistfunc(const coord_t *x, const coord_t *y);
+};
--- a/pgxn/hnsw/test/expected/knn.out
+++ b/pgxn/hnsw/test/expected/knn.out
@@ -0,0 +1,28 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val real[]);
+INSERT INTO t (val) VALUES ('{0,0,0}'), ('{1,2,3}'), ('{1,1,1}'), (NULL);
+CREATE INDEX ON t USING hnsw (val) WITH (maxelements = 10, dims=3, m=3);
+INSERT INTO t (val) VALUES (array[1,2,4]);
+explain SELECT * FROM t ORDER BY val <-> array[3,3,3];
+                             QUERY PLAN                             
+--------------------------------------------------------------------
+ Index Scan using t_val_idx on t  (cost=4.02..8.06 rows=3 width=36)
+   Order By: (val <-> '{3,3,3}'::real[])
+(2 rows)
+
+SELECT * FROM t ORDER BY val <-> array[3,3,3];
+   val   
+---------
+ {1,2,3}
+ {1,2,4}
+ {1,1,1}
+ {0,0,0}
+(4 rows)
+
+SELECT COUNT(*) FROM t;
+ count 
+-------
+     5
+(1 row)
+
+DROP TABLE t;
--- a/pgxn/hnsw/test/sql/knn.sql
+++ b/pgxn/hnsw/test/sql/knn.sql
@@ -0,0 +1,13 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val real[]);
+INSERT INTO t (val) VALUES ('{0,0,0}'), ('{1,2,3}'), ('{1,1,1}'), (NULL);
+CREATE INDEX ON t USING hnsw (val) WITH (maxelements = 10, dims=3, m=3);
+
+INSERT INTO t (val) VALUES (array[1,2,4]);
+
+explain SELECT * FROM t ORDER BY val <-> array[3,3,3];
+SELECT * FROM t ORDER BY val <-> array[3,3,3];
+SELECT COUNT(*) FROM t;
+
+DROP TABLE t;