MB-59575: Handling memory mapped content more responsibly (#241)

Requires:
- blevesearch/faiss#17
- blevesearch/go-faiss#26 (Upgrade to blevesearch/[email protected])

---------

Co-authored-by: Abhinav Dangeti <[email protected]>

Author: 2 people

faiss_vector_cache.go

@@ -118,7 +118,8 @@ func (vc *vectorIndexCache) createAndCache(fieldID uint16, mem []byte, except *r
 	indexSize, n := binary.Uvarint(mem[pos : pos+binary.MaxVarintLen64])
 	pos += n
 
-	index, err = faiss.ReadIndexFromBuffer(mem[pos:pos+int(indexSize)], faiss.IOFlagReadOnly)
+	index, err = faiss.ReadIndexFromBuffer(mem[pos:pos+int(indexSize)],
+		faiss.IOFlagReadMmap|faiss.IOFlagSkipPrefetch)
 	if err != nil {
 		return nil, nil, nil, err
 	}

go.mod

@@ -5,7 +5,7 @@ go 1.20
 require (
 	github.com/RoaringBitmap/roaring v1.9.3
 	github.com/blevesearch/bleve_index_api v1.1.7
-	github.com/blevesearch/go-faiss v1.0.15
+	github.com/blevesearch/go-faiss v1.0.16
 	github.com/blevesearch/mmap-go v1.0.4
 	github.com/blevesearch/scorch_segment_api/v2 v2.2.12
 	github.com/blevesearch/vellum v1.0.10

go.sum

@@ -4,8 +4,8 @@ github.com/bits-and-blooms/bitset v1.12.0 h1:U/q1fAF7xXRhFCrhROzIfffYnu+dlS38vCZ
 github.com/bits-and-blooms/bitset v1.12.0/go.mod h1:7hO7Gc7Pp1vODcmWvKMRA9BNmbv6a/7QIWpPxHddWR8=
 github.com/blevesearch/bleve_index_api v1.1.7 h1:KvvKVn/9+dKCEJHL6flDH3NCiXeKnK3P85e4EOgVjdM=
 github.com/blevesearch/bleve_index_api v1.1.7/go.mod h1:PbcwjIcRmjhGbkS/lJCpfgVSMROV6TRubGGAODaK1W8=
-github.com/blevesearch/go-faiss v1.0.15 h1:aBrj6fwLuY8CkhECFbvkc4qhLTkrYI84QoEaw9z1jMI=
-github.com/blevesearch/go-faiss v1.0.15/go.mod h1:jrxHrbl42X/RnDPI+wBoZU8joxxuRwedrxqswQ3xfU8=
+github.com/blevesearch/go-faiss v1.0.16 h1:lfzXzzjO1mAf15MRiRY5yz6KVGr02CyRrr7m0z70Ih8=
+github.com/blevesearch/go-faiss v1.0.16/go.mod h1:jrxHrbl42X/RnDPI+wBoZU8joxxuRwedrxqswQ3xfU8=
 github.com/blevesearch/mmap-go v1.0.4 h1:OVhDhT5B/M1HNPpYPBKIEJaD0F3Si+CrEKULGCDPWmc=
 github.com/blevesearch/mmap-go v1.0.4/go.mod h1:EWmEAOmdAS9z/pi/+Toxu99DnsbhG1TIxUoRmJw/pSs=
 github.com/blevesearch/scorch_segment_api/v2 v2.2.12 h1:LuBJ4AfKlMsyi/8xFhxIArEnrlAcI3qGTvzwCTdfKgA=

section_faiss_vector_index.go

@@ -286,7 +286,8 @@ func (v *vectorIndexOpaque) mergeAndWriteVectorIndexes(sbs []*SegmentBase,
 		}
 		// read the index bytes. todo: parallelize this
 		indexBytes := segBase.mem[indexes[segI].startOffset : indexes[segI].startOffset+int(indexes[segI].indexSize)]
-		index, err := faiss.ReadIndexFromBuffer(indexBytes, faiss.IOFlagReadOnly)
+		index, err := faiss.ReadIndexFromBuffer(indexBytes,
+			faiss.IOFlagReadMmap|faiss.IOFlagSkipPrefetch)
 		if err != nil {
 			freeReconstructedIndexes(vecIndexes)
 			return err