Compare commits

..

4 Commits

Author SHA1 Message Date
Andrew Kane
310a880186 Use copy [skip ci] 2023-09-12 23:03:59 -07:00
Andrew Kane
3c2a3db8b2 Free datum [skip ci] 2023-09-12 23:01:23 -07:00
Andrew Kane
d57a34b25c Use datum for HNSW 2023-09-12 22:18:19 -07:00
Andrew Kane
9ac825d14e Use datum for lists 2023-09-12 21:13:52 -07:00
21 changed files with 190 additions and 284 deletions

View File

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

View File

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

View File

@@ -1,5 +1,5 @@
EXTENSION = vector
EXTVERSION = 0.5.1
EXTVERSION = 0.5.0
MODULE_big = vector
DATA = $(wildcard sql/*--*.sql)

View File

@@ -1,5 +1,5 @@
EXTENSION = vector
EXTVERSION = 0.5.1
EXTVERSION = 0.5.0
OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
HEADERS = src\vector.h

View File

@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
```sh
cd /tmp
git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
cd pgvector
make
make install # may need sudo
@@ -509,7 +509,7 @@ Then use `nmake` to build:
```cmd
set "PGROOT=C:\Program Files\PostgreSQL\15"
git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
cd pgvector
nmake /F Makefile.win
nmake /F Makefile.win install
@@ -530,7 +530,7 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
You can also build the image manually:
```sh
git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
cd pgvector
docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
```

View File

@@ -1,2 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit

View File

@@ -33,12 +33,6 @@ HnswInit(void)
HNSW_DEFAULT_EF_CONSTRUCTION, HNSW_MIN_EF_CONSTRUCTION, HNSW_MAX_EF_CONSTRUCTION
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
);
add_int_reloption(hnsw_relopt_kind, "dimensions", "Number of dimensions",
HNSW_DEFAULT_DIMENSIONS, HNSW_MIN_DIMENSIONS, HNSW_MAX_DIMENSIONS
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
);
@@ -131,7 +125,6 @@ hnswoptions(Datum reloptions, bool validate)
static const relopt_parse_elt tab[] = {
{"m", RELOPT_TYPE_INT, offsetof(HnswOptions, m)},
{"ef_construction", RELOPT_TYPE_INT, offsetof(HnswOptions, efConstruction)},
{"dimensions", RELOPT_TYPE_INT, offsetof(HnswOptions, dimensions)},
};
#if PG_VERSION_NUM >= 130000

View File

@@ -42,9 +42,6 @@
#define HNSW_DEFAULT_EF_SEARCH 40
#define HNSW_MIN_EF_SEARCH 1
#define HNSW_MAX_EF_SEARCH 1000
#define HNSW_DEFAULT_DIMENSIONS -1
#define HNSW_MIN_DIMENSIONS 1
#define HNSW_MAX_DIMENSIONS HNSW_MAX_DIM
/* Tuple types */
#define HNSW_ELEMENT_TUPLE_TYPE 1
@@ -60,8 +57,6 @@
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
#define PROGRESS_HNSW_PHASE_LOAD 2
#define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
#define HNSW_ELEMENT_TUPLE_SIZE(_datum) MAXALIGN(offsetof(HnswElementTupleData, value) + VARSIZE_ANY(_datum))
#define HNSW_NEIGHBOR_TUPLE_SIZE(level, m) MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
@@ -101,6 +96,7 @@ typedef struct HnswElementData
List *heaptids;
uint8 level;
uint8 deleted;
bool loaded;
HnswNeighborArray *neighbors;
BlockNumber blkno;
OffsetNumber offno;
@@ -115,13 +111,11 @@ typedef struct HnswCandidate
{
HnswElement element;
float distance;
bool closer;
} HnswCandidate;
typedef struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate *items;
} HnswNeighborArray;
@@ -137,7 +131,6 @@ typedef struct HnswOptions
int32 vl_len_; /* varlena header (do not touch directly!) */
int m; /* number of connections */
int efConstruction; /* size of dynamic candidate list */
int dimensions;
} HnswOptions;
typedef struct HnswBuildState
@@ -266,7 +259,6 @@ typedef struct HnswVacuumState
/* Methods */
int HnswGetM(Relation index);
int HnswGetEfConstruction(Relation index);
int HnswGetDimensions(Relation index);
FmgrInfo *HnswOptionalProcInfo(Relation index, uint16 procnum);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
void HnswCommitBuffer(Buffer buf, GenericXLogState *state);

View File

@@ -82,6 +82,7 @@ HnswBuildAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **
HnswPageGetOpaque(*page)->nextblkno = BufferGetBlockNumber(newbuf);
/* Commit */
MarkBufferDirty(*buf);
GenericXLogFinish(*state);
UnlockReleaseBuffer(*buf);
@@ -116,11 +117,11 @@ CreateElementPages(HnswBuildState * buildstate)
ListCell *lc;
/* Calculate sizes */
maxSize = HNSW_MAX_SIZE;
maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
/* Allocate once */
etup = palloc0(BLCKSZ);
ntup = palloc0(BLCKSZ);
etup = palloc0(maxSize);
ntup = palloc0(maxSize);
/* Prepare first page */
buf = HnswNewBuffer(index, forkNum);
@@ -178,6 +179,7 @@ CreateElementPages(HnswBuildState * buildstate)
insertPage = BufferGetBlockNumber(buf);
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
@@ -225,6 +227,7 @@ CreateNeighborPages(HnswBuildState * buildstate)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -375,13 +378,11 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
if (inserted)
{
element->value = datumCopy(element->value, false, -1);
element->loaded = true;
buildstate->elements = lappend(buildstate->elements, element);
}
else
{
element->value = PointerGetDatum(NULL);
HnswFreeElement(element);
}
/* Reset memory context */
MemoryContextReset(buildstate->tmpCtx);
@@ -399,7 +400,6 @@ HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
elementSize += sizeof(ItemPointerData);
/* TODO Handle non-vector types */
elementSize += VECTOR_SIZE(dimensions);
return (maintenance_work_mem * 1024L) / elementSize;
}
@@ -417,10 +417,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->m = HnswGetM(index);
buildstate->efConstruction = HnswGetEfConstruction(index);
buildstate->dimensions = HnswGetDimensions(index);
if (buildstate->dimensions < 0)
buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)

View File

@@ -134,7 +134,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
etupSize = HNSW_ELEMENT_TUPLE_SIZE(e->value);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
maxSize = HNSW_MAX_SIZE;
maxSize = BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData));
minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
/* Prepare element tuple */
@@ -201,6 +201,8 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
@@ -267,6 +269,9 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
}
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
if (nbuf != buf)
@@ -385,6 +390,7 @@ HnswUpdateNeighborPages(Relation index, FmgrInfo *procinfo, Oid collation, HnswE
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
@@ -438,6 +444,7 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);

View File

@@ -113,6 +113,12 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
scan->opaque = so;
/*
* Get a shared lock. This allows vacuum to ensure no in-flight scans
* before marking tuples as deleted.
*/
LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
return scan;
}
@@ -160,25 +166,11 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan hnsw index without order");
/* Requires MVCC-compliant snapshot as not able to maintain a pin */
/* https://www.postgresql.org/docs/current/index-locking.html */
if (!IsMVCCSnapshot(scan->xs_snapshot))
elog(ERROR, "non-MVCC snapshots are not supported with hnsw");
/* Get scan value */
value = GetScanValue(scan);
/*
* Get a shared lock. This allows vacuum to ensure no in-flight scans
* before marking tuples as deleted.
*/
LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
so->w = GetScanItems(scan, value);
/* Release shared lock */
UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
so->first = false;
}
@@ -206,6 +198,15 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
scan->xs_ctup.t_self = *heaptid;
#endif
/*
* Typically, an index scan must maintain a pin on the index page
* holding the item last returned by amgettuple. However, this is not
* needed with the current vacuum strategy, which ensures scans do not
* visit tuples in danger of being marked as deleted.
*
* https://www.postgresql.org/docs/current/index-locking.html
*/
scan->xs_recheckorderby = false;
return true;
}
@@ -222,6 +223,9 @@ hnswendscan(IndexScanDesc scan)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
/* Release shared lock */
UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
MemoryContextDelete(so->tmpCtx);
pfree(so);

View File

@@ -35,20 +35,6 @@ HnswGetEfConstruction(Relation index)
return HNSW_DEFAULT_EF_CONSTRUCTION;
}
/*
* Get the number of dimensions in the index
*/
int
HnswGetDimensions(Relation index)
{
HnswOptions *opts = (HnswOptions *) index->rd_options;
if (opts)
return opts->dimensions;
return HNSW_DEFAULT_DIMENSIONS;
}
/*
* Get proc
*/
@@ -132,6 +118,7 @@ HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState *
void
HnswCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -154,7 +141,6 @@ HnswInitNeighbors(HnswElement element, int m)
a = &element->neighbors[lc];
a->length = 0;
a->items = palloc(sizeof(HnswCandidate) * lm);
a->closerSet = false;
}
}
@@ -202,7 +188,7 @@ HnswFreeElement(HnswElement element)
{
HnswFreeNeighbors(element);
list_free_deep(element->heaptids);
if (DatumGetPointer(element->value))
if (element->loaded)
pfree(DatumGetPointer(element->value));
pfree(element);
}
@@ -230,7 +216,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
element->blkno = blkno;
element->offno = offno;
element->neighbors = NULL;
element->value = PointerGetDatum(NULL);
element->loaded = false;
return element;
}
@@ -466,6 +452,7 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
Datum value = PointerGetDatum(&etup->value);
element->value = datumCopy(value, false, -1);
element->loaded = true;
}
}
@@ -710,34 +697,6 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
return w;
}
/*
* Compare candidate distances
*/
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistances(const ListCell *a, const ListCell *b)
#else
CompareCandidateDistances(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
if (hca->distance < hcb->distance)
return 1;
if (hca->distance > hcb->distance)
return -1;
if (hca->element < hcb->element)
return 1;
if (hca->element > hcb->element)
return -1;
return 0;
}
/*
* Calculate the distance between elements
*/
@@ -794,77 +753,33 @@ 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(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswCandidate * *pruned)
{
List *r = NIL;
List *w = list_copy(c);
pairingheap *wd;
bool mustCalculate = !e2->neighbors[lc].closerSet;
List *added = NIL;
bool removedAny = false;
if (list_length(w) <= m)
return w;
wd = pairingheap_allocate(CompareNearestCandidates, NULL);
/* Ensure order of candidates is deterministic for closer caching */
if (sortCandidates)
list_sort(w, CompareCandidateDistances);
while (list_length(w) > 0 && list_length(r) < m)
{
/* Assumes w is already ordered desc */
HnswCandidate *e = llast(w);
bool closer;
w = list_delete_last(w);
/* Use previous state of r and wd to skip work when possible */
if (mustCalculate)
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
else if (list_length(added) > 0)
{
/*
* If the current candidate was closer, we only need to compare it
* with the other candidates that we have added.
*/
if (e->closer)
{
e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
closer = CheckElementCloser(e, r, lc, procinfo, collation);
if (!e->closer)
removedAny = true;
}
else
{
/*
* If we have removed any candidates from closer, a candidate
* that was not closer earlier might now be.
*/
if (removedAny)
{
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
if (e->closer)
added = lappend(added, e);
}
}
}
else if (e == newCandidate)
{
e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
if (e->closer)
added = lappend(added, e);
}
if (e->closer)
if (closer)
r = lappend(r, e);
else
pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
}
/* Cached value can only be used in future if sorted deterministically */
e2->neighbors[lc].closerSet = sortCandidates;
/* Keep pruned connections */
while (!pairingheap_is_empty(wd) && list_length(r) < m)
r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -918,6 +833,28 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
}
/*
* Compare candidate distances
*/
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistances(const ListCell *a, const ListCell *b)
#else
CompareCandidateDistances(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
if (hca->distance < hcb->distance)
return 1;
if (hca->distance > hcb->distance)
return -1;
return 0;
}
/*
* Update connections
*/
@@ -953,7 +890,7 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
{
HnswCandidate *hc3 = &currentNeighbors->items[i];
if (!DatumGetPointer(hc3->element->value))
if (!hc3->element->loaded)
HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
else
hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -971,12 +908,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
{
List *c = NIL;
/* Add candidates */
/* Add and sort candidates */
for (int i = 0; i < currentNeighbors->length; i++)
c = lappend(c, &currentNeighbors->items[i]);
c = lappend(c, &hc2);
list_sort(c, CompareCandidateDistances);
SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
SelectNeighbors(c, m, lc, procinfo, collation, &pruned);
/* Should not happen */
if (pruned == NULL)
@@ -1075,12 +1013,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
else
lw = w;
/*
* Candidates are sorted, but not deterministically. Could set
* sortCandidates to true for in-memory builds to enable closer
* caching, but there does not seem to be a difference in performance.
*/
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL, false);
neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
AddConnections(element, neighbors, lm, lc);

View File

@@ -128,7 +128,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
blkno = HnswPageGetOpaque(page)->nextblkno;
if (updated)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
@@ -226,6 +229,7 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
@@ -547,6 +551,9 @@ MarkDeleted(HnswVacuumState * vacuumstate)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
if (nbuf != buf)
UnlockReleaseBuffer(nbuf);

View File

@@ -506,11 +506,9 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
Buffer buf;
Page page;
GenericXLogState *state;
Size listSize;
IvfflatList list;
listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
list = palloc(listSize);
list = palloc0(BLCKSZ);
buf = IvfflatNewBuffer(index, forkNum);
IvfflatInitRegisterPage(index, &buf, &page, &state);
@@ -518,11 +516,13 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
for (int i = 0; i < lists; i++)
{
OffsetNumber offno;
Datum center = PointerGetDatum(VectorArrayGet(centers, i));
Size listSize = MAXALIGN(IVFFLAT_LIST_SIZE(center));
/* Load list */
list->startPage = InvalidBlockNumber;
list->insertPage = InvalidBlockNumber;
memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
memcpy(&list->center, DatumGetPointer(center), VARSIZE_ANY(center));
/* Ensure free space */
if (PageGetFreeSpace(page) < listSize)

View File

@@ -52,7 +52,7 @@
#define PROGRESS_IVFFLAT_PHASE_ASSIGN 3
#define PROGRESS_IVFFLAT_PHASE_LOAD 4
#define IVFFLAT_LIST_SIZE(_dim) (offsetof(IvfflatListData, center) + VECTOR_SIZE(_dim))
#define IVFFLAT_LIST_SIZE(_datum) (offsetof(IvfflatListData, center) + VARSIZE_ANY(_datum))
#define IvfflatPageGetOpaque(page) ((IvfflatPageOpaque) PageGetSpecialPointer(page))
#define IvfflatPageGetMeta(page) ((IvfflatMetaPageData *) PageGetContents(page))
@@ -229,7 +229,7 @@ typedef struct IvfflatListData
{
BlockNumber startPage;
BlockNumber insertPage;
Vector center;
char center[FLEXIBLE_ARRAY_MEMBER];
} IvfflatListData;
typedef IvfflatListData * IvfflatList;
@@ -246,6 +246,8 @@ typedef struct IvfflatScanOpaqueData
int probes;
int dimensions;
bool first;
Buffer buf;
ItemPointerData heaptid;
/* Sorting */
Tuplesortstate *sortstate;

View File

@@ -99,7 +99,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
/* Get tuple size */
itemsz = MAXALIGN(IndexTupleSize(itup));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)) - sizeof(ItemIdData));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
/* Find a page to insert the item */
for (;;)
@@ -142,6 +142,8 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
IvfflatPageGetOpaque(page)->nextblkno = insertPage;
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */

View File

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

View File

@@ -136,6 +136,7 @@ IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogStat
void
IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -159,6 +160,8 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
IvfflatInitPage(newbuf, newpage);
/* Commit */
MarkBufferDirty(*buf);
MarkBufferDirty(newbuf);
GenericXLogFinish(*state);
/* Unlock */

View File

@@ -107,6 +107,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
{
/* Delete tuples */
PageIndexMultiDelete(page, deletable, ndeletable);
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else

View File

@@ -1,113 +0,0 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
sub test_recall
{
my ($min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Scan/);
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my %actual_set = map { $_ => 1 } @actual_ids;
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
}
}
cmp_ok($correct / $total, ">=", $min, $operator);
}
# Initialize node
$node = get_new_node('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v float4[3]);");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", qq(
CREATE FUNCTION float4_l2_distance(float4[], float4[]) RETURNS float8
AS 'BEGIN RETURN l2_distance(\$1::vector, \$2::vector); END;'
LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION float4_l2_squared_distance(float4[], float4[]) RETURNS float8
AS 'BEGIN RETURN vector_l2_squared_distance(\$1::vector, \$2::vector); END;'
LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR <-> (
LEFTARG = float4[], RIGHTARG = float4[], PROCEDURE = float4_l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR CLASS float4_l2_ops
FOR TYPE float4[] USING hnsw AS
OPERATOR 1 <-> (float4[], float4[]) FOR ORDER BY float_ops,
FUNCTION 1 float4_l2_squared_distance(float4[], float4[]);
));
# Generate queries
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
push(@queries, "{$r1,$r2,$r3}");
}
# Check each index type
my @operators = ("<->");
my @opclasses = ("float4_l2_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
# Add index
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass) WITH (dimensions = 3);");
my $min = $operator eq "<#>" ? 0.80 : 0.99;
test_recall($min, $operator);
}
done_testing();

View File

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