Compare commits

..

3 Commits

Author SHA1 Message Date
Andrew Kane
7132c9b111 Use fabs [skip ci] 2023-05-05 18:34:15 -07:00
Andrew Kane
878e1e6c3a Added query-aware dynamic pruning [skip ci] 2023-05-05 18:13:56 -07:00
Andrew Kane
d885e2bcfa Added FAQ about results [skip ci] 2023-05-02 10:17:31 -07:00
4 changed files with 17 additions and 2 deletions

View File

@@ -312,6 +312,10 @@ Two things you can try are:
1. use dimensionality reduction 1. use dimensionality reduction
2. compile Postgres with a larger block size (`./configure --with-blocksize=32`) and edit the limit in `src/ivfflat.h` 2. compile Postgres with a larger block size (`./configure --with-blocksize=32`) and edit the limit in `src/ivfflat.h`
#### Why am I seeing less results after adding an index?
The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
## Reference ## Reference
### Vector Type ### Vector Type

View File

@@ -419,7 +419,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
/* Target 50 samples per list, with at least 10000 samples */ /* Target 50 samples per list, with at least 10000 samples */
/* The number of samples has a large effect on index build time */ /* The number of samples has a large effect on index build time */
numSamples = buildstate->lists * 200; numSamples = buildstate->lists * 50;
if (numSamples < 10000) if (numSamples < 10000)
numSamples = 10000; numSamples = 10000;

View File

@@ -206,6 +206,7 @@ typedef struct IvfflatScanOpaqueData
/* Lists */ /* Lists */
pairingheap *listQueue; pairingheap *listQueue;
double minDistance;
IvfflatScanList lists[FLEXIBLE_ARRAY_MEMBER]; /* must come last */ IvfflatScanList lists[FLEXIBLE_ARRAY_MEMBER]; /* must come last */
} IvfflatScanOpaqueData; } IvfflatScanOpaqueData;

View File

@@ -60,6 +60,9 @@ GetScanLists(IndexScanDesc scan, Datum value)
/* Use procinfo from the index instead of scan key for performance */ /* Use procinfo from the index instead of scan key for performance */
distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value)); distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
if (distance < so->minDistance)
so->minDistance = distance;
if (listCount < so->probes) if (listCount < so->probes)
{ {
scanlist = &so->lists[listCount]; scanlist = &so->lists[listCount];
@@ -129,7 +132,13 @@ GetScanItems(IndexScanDesc scan, Datum value)
/* Search closest probes lists */ /* Search closest probes lists */
while (!pairingheap_is_empty(so->listQueue)) while (!pairingheap_is_empty(so->listQueue))
{ {
searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage; IvfflatScanList *scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
/* Query-aware dynamic pruning */
if (fabs(scanlist->distance) > 1.5 * fabs(so->minDistance))
continue;
searchPage = scanlist->startPage;
/* Search all entry pages for list */ /* Search all entry pages for list */
while (BlockNumberIsValid(searchPage)) while (BlockNumberIsValid(searchPage))
@@ -252,6 +261,7 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
so->first = true; so->first = true;
pairingheap_reset(so->listQueue); pairingheap_reset(so->listQueue);
so->minDistance = DBL_MAX;
if (keys && scan->numberOfKeys > 0) if (keys && scan->numberOfKeys > 0)
memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData)); memmove(scan->keyData, keys, scan->numberOfKeys * sizeof(ScanKeyData));