You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@lucene.apache.org by GitBox <gi...@apache.org> on 2022/11/02 20:14:29 UTC
[GitHub] [lucene] rmuir commented on a diff in pull request #11860: GITHUB-11830 Better optimize storage for vector connections
rmuir commented on code in PR #11860:
URL: https://github.com/apache/lucene/pull/11860#discussion_r1012255076
##########
lucene/core/src/java/org/apache/lucene/codecs/lucene95/Lucene95HnswVectorsReader.java:
##########
@@ -0,0 +1,522 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.lucene.codecs.lucene95;
+
+import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS;
+
+import java.io.IOException;
+import java.util.Arrays;
+import java.util.HashMap;
+import java.util.Map;
+import org.apache.lucene.codecs.CodecUtil;
+import org.apache.lucene.codecs.KnnVectorsReader;
+import org.apache.lucene.index.*;
+import org.apache.lucene.search.ScoreDoc;
+import org.apache.lucene.search.TopDocs;
+import org.apache.lucene.search.TotalHits;
+import org.apache.lucene.store.ChecksumIndexInput;
+import org.apache.lucene.store.DataInput;
+import org.apache.lucene.store.IndexInput;
+import org.apache.lucene.util.Bits;
+import org.apache.lucene.util.IOUtils;
+import org.apache.lucene.util.RamUsageEstimator;
+import org.apache.lucene.util.hnsw.HnswGraph;
+import org.apache.lucene.util.hnsw.HnswGraphSearcher;
+import org.apache.lucene.util.hnsw.NeighborQueue;
+import org.apache.lucene.util.packed.DirectMonotonicReader;
+import org.apache.lucene.util.packed.PackedInts;
+
+/**
+ * Reads vectors from the index segments along with index data structures supporting KNN search.
+ *
+ * @lucene.experimental
+ */
+public final class Lucene95HnswVectorsReader extends KnnVectorsReader {
+
+ private final FieldInfos fieldInfos;
+ private final Map<String, FieldEntry> fields = new HashMap<>();
+ private final IndexInput vectorData;
+ private final IndexInput vectorIndex;
+
+ Lucene95HnswVectorsReader(SegmentReadState state) throws IOException {
+ this.fieldInfos = state.fieldInfos;
+ int versionMeta = readMetadata(state);
+ boolean success = false;
+ try {
+ vectorData =
+ openDataInput(
+ state,
+ versionMeta,
+ Lucene95HnswVectorsFormat.VECTOR_DATA_EXTENSION,
+ Lucene95HnswVectorsFormat.VECTOR_DATA_CODEC_NAME);
+ vectorIndex =
+ openDataInput(
+ state,
+ versionMeta,
+ Lucene95HnswVectorsFormat.VECTOR_INDEX_EXTENSION,
+ Lucene95HnswVectorsFormat.VECTOR_INDEX_CODEC_NAME);
+ success = true;
+ } finally {
+ if (success == false) {
+ IOUtils.closeWhileHandlingException(this);
+ }
+ }
+ }
+
+ private int readMetadata(SegmentReadState state) throws IOException {
+ String metaFileName =
+ IndexFileNames.segmentFileName(
+ state.segmentInfo.name, state.segmentSuffix, Lucene95HnswVectorsFormat.META_EXTENSION);
+ int versionMeta = -1;
+ try (ChecksumIndexInput meta = state.directory.openChecksumInput(metaFileName, state.context)) {
+ Throwable priorE = null;
+ try {
+ versionMeta =
+ CodecUtil.checkIndexHeader(
+ meta,
+ Lucene95HnswVectorsFormat.META_CODEC_NAME,
+ Lucene95HnswVectorsFormat.VERSION_START,
+ Lucene95HnswVectorsFormat.VERSION_CURRENT,
+ state.segmentInfo.getId(),
+ state.segmentSuffix);
+ readFields(meta, state.fieldInfos);
+ } catch (Throwable exception) {
+ priorE = exception;
+ } finally {
+ CodecUtil.checkFooter(meta, priorE);
+ }
+ }
+ return versionMeta;
+ }
+
+ private static IndexInput openDataInput(
+ SegmentReadState state, int versionMeta, String fileExtension, String codecName)
+ throws IOException {
+ String fileName =
+ IndexFileNames.segmentFileName(state.segmentInfo.name, state.segmentSuffix, fileExtension);
+ IndexInput in = state.directory.openInput(fileName, state.context);
+ boolean success = false;
+ try {
+ int versionVectorData =
+ CodecUtil.checkIndexHeader(
+ in,
+ codecName,
+ Lucene95HnswVectorsFormat.VERSION_START,
+ Lucene95HnswVectorsFormat.VERSION_CURRENT,
+ state.segmentInfo.getId(),
+ state.segmentSuffix);
+ if (versionMeta != versionVectorData) {
+ throw new CorruptIndexException(
+ "Format versions mismatch: meta="
+ + versionMeta
+ + ", "
+ + codecName
+ + "="
+ + versionVectorData,
+ in);
+ }
+ CodecUtil.retrieveChecksum(in);
+ success = true;
+ return in;
+ } finally {
+ if (success == false) {
+ IOUtils.closeWhileHandlingException(in);
+ }
+ }
+ }
+
+ private void readFields(ChecksumIndexInput meta, FieldInfos infos) throws IOException {
+ for (int fieldNumber = meta.readInt(); fieldNumber != -1; fieldNumber = meta.readInt()) {
+ FieldInfo info = infos.fieldInfo(fieldNumber);
+ if (info == null) {
+ throw new CorruptIndexException("Invalid field number: " + fieldNumber, meta);
+ }
+ FieldEntry fieldEntry = readField(meta);
+ validateFieldEntry(info, fieldEntry);
+ fields.put(info.name, fieldEntry);
+ }
+ }
+
+ private void validateFieldEntry(FieldInfo info, FieldEntry fieldEntry) {
+ int dimension = info.getVectorDimension();
+ if (dimension != fieldEntry.dimension) {
+ throw new IllegalStateException(
+ "Inconsistent vector dimension for field=\""
+ + info.name
+ + "\"; "
+ + dimension
+ + " != "
+ + fieldEntry.dimension);
+ }
+
+ int byteSize =
+ switch (info.getVectorEncoding()) {
+ case BYTE -> Byte.BYTES;
+ case FLOAT32 -> Float.BYTES;
+ };
+ long vectorBytes = Math.multiplyExact((long) dimension, byteSize);
+ long numBytes = Math.multiplyExact(vectorBytes, fieldEntry.size);
+ if (numBytes != fieldEntry.vectorDataLength) {
+ throw new IllegalStateException(
+ "Vector data length "
+ + fieldEntry.vectorDataLength
+ + " not matching size="
+ + fieldEntry.size
+ + " * dim="
+ + dimension
+ + " * byteSize="
+ + byteSize
+ + " = "
+ + numBytes);
+ }
+ }
+
+ private VectorSimilarityFunction readSimilarityFunction(DataInput input) throws IOException {
+ int similarityFunctionId = input.readInt();
+ if (similarityFunctionId < 0
+ || similarityFunctionId >= VectorSimilarityFunction.values().length) {
+ throw new CorruptIndexException(
+ "Invalid similarity function id: " + similarityFunctionId, input);
+ }
+ return VectorSimilarityFunction.values()[similarityFunctionId];
+ }
+
+ private VectorEncoding readVectorEncoding(DataInput input) throws IOException {
+ int encodingId = input.readInt();
+ if (encodingId < 0 || encodingId >= VectorEncoding.values().length) {
+ throw new CorruptIndexException("Invalid vector encoding id: " + encodingId, input);
+ }
+ return VectorEncoding.values()[encodingId];
+ }
+
+ private FieldEntry readField(IndexInput input) throws IOException {
+ VectorEncoding vectorEncoding = readVectorEncoding(input);
+ VectorSimilarityFunction similarityFunction = readSimilarityFunction(input);
+ return new FieldEntry(input, vectorEncoding, similarityFunction);
+ }
+
+ @Override
+ public long ramBytesUsed() {
+ long totalBytes = RamUsageEstimator.shallowSizeOfInstance(Lucene95HnswVectorsFormat.class);
+ totalBytes +=
+ RamUsageEstimator.sizeOfMap(
+ fields, RamUsageEstimator.shallowSizeOfInstance(FieldEntry.class));
+ return totalBytes;
+ }
+
+ @Override
+ public void checkIntegrity() throws IOException {
+ CodecUtil.checksumEntireFile(vectorData);
+ CodecUtil.checksumEntireFile(vectorIndex);
+ }
+
+ @Override
+ public VectorValues getVectorValues(String field) throws IOException {
+ FieldEntry fieldEntry = fields.get(field);
+ VectorValues values = OffHeapVectorValues.load(fieldEntry, vectorData);
+ if (fieldEntry.vectorEncoding == VectorEncoding.BYTE) {
+ return new ExpandingVectorValues(values);
+ } else {
+ return values;
+ }
+ }
+
+ @Override
+ public TopDocs search(String field, float[] target, int k, Bits acceptDocs, int visitedLimit)
+ throws IOException {
+ FieldEntry fieldEntry = fields.get(field);
+
+ if (fieldEntry.size() == 0) {
+ return new TopDocs(new TotalHits(0, TotalHits.Relation.EQUAL_TO), new ScoreDoc[0]);
+ }
+
+ // bound k by total number of vectors to prevent oversizing data structures
+ k = Math.min(k, fieldEntry.size());
+ OffHeapVectorValues vectorValues = OffHeapVectorValues.load(fieldEntry, vectorData);
+
+ NeighborQueue results =
+ HnswGraphSearcher.search(
+ target,
+ k,
+ vectorValues,
+ fieldEntry.vectorEncoding,
+ fieldEntry.similarityFunction,
+ getGraph(fieldEntry),
+ vectorValues.getAcceptOrds(acceptDocs),
+ visitedLimit);
+
+ int i = 0;
+ ScoreDoc[] scoreDocs = new ScoreDoc[Math.min(results.size(), k)];
+ while (results.size() > 0) {
+ int node = results.topNode();
+ float score = results.topScore();
+ results.pop();
+ scoreDocs[scoreDocs.length - ++i] = new ScoreDoc(vectorValues.ordToDoc(node), score);
+ }
+
+ TotalHits.Relation relation =
+ results.incomplete()
+ ? TotalHits.Relation.GREATER_THAN_OR_EQUAL_TO
+ : TotalHits.Relation.EQUAL_TO;
+ return new TopDocs(new TotalHits(results.visitedCount(), relation), scoreDocs);
+ }
+
+ /** Get knn graph values; used for testing */
+ public HnswGraph getGraph(String field) throws IOException {
+ FieldInfo info = fieldInfos.fieldInfo(field);
+ if (info == null) {
+ throw new IllegalArgumentException("No such field '" + field + "'");
+ }
+ FieldEntry entry = fields.get(field);
+ if (entry != null && entry.vectorIndexLength > 0) {
+ return getGraph(entry);
+ } else {
+ return HnswGraph.EMPTY;
+ }
+ }
+
+ private HnswGraph getGraph(FieldEntry entry) throws IOException {
+ IndexInput bytesSlice =
+ vectorIndex.slice("graph-data", entry.vectorIndexOffset, entry.vectorIndexLength);
+ return new OffHeapHnswGraph(entry, bytesSlice);
+ }
+
+ @Override
+ public void close() throws IOException {
+ IOUtils.close(vectorData, vectorIndex);
+ }
+
+ static class FieldEntry {
+
+ final VectorSimilarityFunction similarityFunction;
+ final VectorEncoding vectorEncoding;
+ final long vectorDataOffset;
+ final long vectorDataLength;
+ final long vectorIndexOffset;
+ final long vectorIndexLength;
+ final int M;
+ final int numLevels;
+ final int dimension;
+ final int size;
+ final int[][] nodesByLevel;
+ // for each level the start offsets in vectorIndex file from where to read neighbours
+ final long[] graphOffsetsByLevel;
+ // the packed ints version used to store the graph connections
+ final int packedIntsVersion;
+
+ // the following four variables used to read docIds encoded by IndexDISI
+ // special values of docsWithFieldOffset are -1 and -2
+ // -1 : dense
+ // -2 : empty
+ // other: sparse
+ final long docsWithFieldOffset;
+ final long docsWithFieldLength;
+ final short jumpTableEntryCount;
+ final byte denseRankPower;
+
+ // the following four variables used to read ordToDoc encoded by DirectMonotonicWriter
+ // note that only spare case needs to store ordToDoc
+ final long addressesOffset;
+ final int blockShift;
+ final DirectMonotonicReader.Meta meta;
+ final long addressesLength;
+
+ FieldEntry(
+ IndexInput input,
+ VectorEncoding vectorEncoding,
+ VectorSimilarityFunction similarityFunction)
+ throws IOException {
+ this.similarityFunction = similarityFunction;
+ this.vectorEncoding = vectorEncoding;
+ vectorDataOffset = input.readVLong();
+ vectorDataLength = input.readVLong();
+ vectorIndexOffset = input.readVLong();
+ vectorIndexLength = input.readVLong();
+ dimension = input.readInt();
+ size = input.readInt();
+
+ docsWithFieldOffset = input.readLong();
+ docsWithFieldLength = input.readLong();
+ jumpTableEntryCount = input.readShort();
+ denseRankPower = input.readByte();
+
+ // dense or empty
+ if (docsWithFieldOffset == -1 || docsWithFieldOffset == -2) {
+ addressesOffset = 0;
+ blockShift = 0;
+ meta = null;
+ addressesLength = 0;
+ } else {
+ // sparse
+ addressesOffset = input.readLong();
+ blockShift = input.readVInt();
+ meta = DirectMonotonicReader.loadMeta(input, size, blockShift);
+ addressesLength = input.readLong();
+ }
+
+ // read nodes by level
+ M = input.readInt();
+ packedIntsVersion = input.readVInt();
+ numLevels = input.readInt();
+ nodesByLevel = new int[numLevels][];
+ for (int level = 0; level < numLevels; level++) {
+ int numNodesOnLevel = input.readInt();
+ if (level == 0) {
+ // we don't store nodes for level 0th, as this level contains all nodes
+ assert numNodesOnLevel == size;
+ nodesByLevel[0] = null;
+ } else {
+ nodesByLevel[level] = new int[numNodesOnLevel];
+ for (int i = 0; i < numNodesOnLevel; i++) {
+ nodesByLevel[level][i] = input.readInt();
+ }
+ }
+ }
+
+ // calculate for each level the start offsets in vectorIndex file from where to read
+ // neighbours
+ graphOffsetsByLevel = new long[numLevels];
+ final int packedBitsRequired = PackedInts.bitsRequired(size);
+ for (int level = 0; level < numLevels; level++) {
+ if (level == 0) {
+ graphOffsetsByLevel[level] = 0;
+ } else if (level == 1) {
+ graphOffsetsByLevel[level] =
+ (PackedInts.Format.PACKED.byteCount(packedIntsVersion, M * 2, packedBitsRequired)
+ + Integer.BYTES)
+ * size;
Review Comment:
should we use addExact/multiplyExact here? it's a one-time initialization so we can add some safety around overflow in these parts, especially since there is multiplication happening.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org