benwtrent commented on code in PR #12582: URL: https://github.com/apache/lucene/pull/12582#discussion_r1344010395
########## lucene/core/src/java/org/apache/lucene/codecs/lucene99/Lucene99HnswVectorsWriter.java: ########## @@ -0,0 +1,1170 @@ +/* + * 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.lucene99; + +import static org.apache.lucene.codecs.lucene99.Lucene99HnswVectorsFormat.DIRECT_MONOTONIC_BLOCK_SHIFT; +import static org.apache.lucene.codecs.lucene99.Lucene99ScalarQuantizedVectorsFormat.calculateDefaultQuantile; +import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS; + +import java.io.IOException; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Collections; +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import org.apache.lucene.codecs.CodecUtil; +import org.apache.lucene.codecs.HnswGraphProvider; +import org.apache.lucene.codecs.KnnFieldVectorsWriter; +import org.apache.lucene.codecs.KnnVectorsReader; +import org.apache.lucene.codecs.KnnVectorsWriter; +import org.apache.lucene.codecs.lucene90.IndexedDISI; +import org.apache.lucene.codecs.perfield.PerFieldKnnVectorsFormat; +import org.apache.lucene.index.ByteVectorValues; +import org.apache.lucene.index.DocsWithFieldSet; +import org.apache.lucene.index.FieldInfo; +import org.apache.lucene.index.FloatVectorValues; +import org.apache.lucene.index.IndexFileNames; +import org.apache.lucene.index.MergeState; +import org.apache.lucene.index.SegmentWriteState; +import org.apache.lucene.index.Sorter; +import org.apache.lucene.index.VectorEncoding; +import org.apache.lucene.search.DocIdSetIterator; +import org.apache.lucene.store.IndexInput; +import org.apache.lucene.store.IndexOutput; +import org.apache.lucene.util.ArrayUtil; +import org.apache.lucene.util.Bits; +import org.apache.lucene.util.IOUtils; +import org.apache.lucene.util.InfoStream; +import org.apache.lucene.util.RamUsageEstimator; +import org.apache.lucene.util.hnsw.HnswGraph; +import org.apache.lucene.util.hnsw.HnswGraph.NodesIterator; +import org.apache.lucene.util.hnsw.HnswGraphBuilder; +import org.apache.lucene.util.hnsw.NeighborArray; +import org.apache.lucene.util.hnsw.OnHeapHnswGraph; +import org.apache.lucene.util.hnsw.RandomAccessVectorValues; +import org.apache.lucene.util.hnsw.RandomVectorScorerSupplier; +import org.apache.lucene.util.packed.DirectMonotonicWriter; + +/** + * Writes vector values and knn graphs to index segments. + * + * @lucene.experimental + */ +public final class Lucene99HnswVectorsWriter extends KnnVectorsWriter { + + private final SegmentWriteState segmentWriteState; + private final IndexOutput meta, vectorData, quantizedVectorData, vectorIndex; + private final int M; + private final int beamWidth; + private final Lucene99ScalarQuantizedVectorsWriter quantizedVectorsWriter; + + private final List<FieldWriter<?>> fields = new ArrayList<>(); + private boolean finished; + + Lucene99HnswVectorsWriter( + SegmentWriteState state, + int M, + int beamWidth, + Lucene99ScalarQuantizedVectorsFormat quantizedVectorsFormat) + throws IOException { + this.M = M; + this.beamWidth = beamWidth; + segmentWriteState = state; + String metaFileName = + IndexFileNames.segmentFileName( + state.segmentInfo.name, state.segmentSuffix, Lucene99HnswVectorsFormat.META_EXTENSION); + + String vectorDataFileName = + IndexFileNames.segmentFileName( + state.segmentInfo.name, + state.segmentSuffix, + Lucene99HnswVectorsFormat.VECTOR_DATA_EXTENSION); + + String indexDataFileName = + IndexFileNames.segmentFileName( + state.segmentInfo.name, + state.segmentSuffix, + Lucene99HnswVectorsFormat.VECTOR_INDEX_EXTENSION); + + final String quantizedVectorDataFileName = + quantizedVectorsFormat != null + ? IndexFileNames.segmentFileName( + state.segmentInfo.name, + state.segmentSuffix, + Lucene99ScalarQuantizedVectorsFormat.QUANTIZED_VECTOR_DATA_EXTENSION) + : null; + boolean success = false; + try { + meta = state.directory.createOutput(metaFileName, state.context); + vectorData = state.directory.createOutput(vectorDataFileName, state.context); + vectorIndex = state.directory.createOutput(indexDataFileName, state.context); + + CodecUtil.writeIndexHeader( + meta, + Lucene99HnswVectorsFormat.META_CODEC_NAME, + Lucene99HnswVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + CodecUtil.writeIndexHeader( + vectorData, + Lucene99HnswVectorsFormat.VECTOR_DATA_CODEC_NAME, + Lucene99HnswVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + CodecUtil.writeIndexHeader( + vectorIndex, + Lucene99HnswVectorsFormat.VECTOR_INDEX_CODEC_NAME, + Lucene99HnswVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + if (quantizedVectorDataFileName != null) { + quantizedVectorData = + state.directory.createOutput(quantizedVectorDataFileName, state.context); + CodecUtil.writeIndexHeader( + quantizedVectorData, + Lucene99ScalarQuantizedVectorsFormat.QUANTIZED_VECTOR_DATA_CODEC_NAME, + Lucene99ScalarQuantizedVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + quantizedVectorsWriter = + new Lucene99ScalarQuantizedVectorsWriter( + quantizedVectorData, quantizedVectorsFormat.quantile); + } else { + quantizedVectorData = null; + quantizedVectorsWriter = null; + } + success = true; + } finally { + if (success == false) { + IOUtils.closeWhileHandlingException(this); + } + } + } + + @Override + public KnnFieldVectorsWriter<?> addField(FieldInfo fieldInfo) throws IOException { + Lucene99ScalarQuantizedVectorsWriter.QuantizationVectorWriter quantizedVectorFieldWriter = null; + // Quantization only supports FLOAT32 for now + if (quantizedVectorsWriter != null + && fieldInfo.getVectorEncoding().equals(VectorEncoding.FLOAT32)) { + quantizedVectorFieldWriter = quantizedVectorsWriter.addField(fieldInfo); + } + FieldWriter<?> newField = + FieldWriter.create( + fieldInfo, M, beamWidth, segmentWriteState.infoStream, quantizedVectorFieldWriter); + fields.add(newField); + return newField; + } + + @Override + public void flush(int maxDoc, Sorter.DocMap sortMap) throws IOException { + for (FieldWriter<?> field : fields) { + long[] quantizedVectorOffsetAndLen = null; + if (field.quantizedWriter != null) { + assert quantizedVectorsWriter != null; + quantizedVectorOffsetAndLen = + quantizedVectorsWriter.flush(sortMap, field.quantizedWriter, field.docsWithField); + } + if (sortMap == null) { + writeField(field, maxDoc, quantizedVectorOffsetAndLen); + } else { + writeSortingField(field, maxDoc, sortMap, quantizedVectorOffsetAndLen); + } + } + } + + @Override + public void finish() throws IOException { + if (finished) { + throw new IllegalStateException("already finished"); + } + finished = true; + if (quantizedVectorsWriter != null) { + quantizedVectorsWriter.finish(); + } + + if (meta != null) { + // write end of fields marker + meta.writeInt(-1); + CodecUtil.writeFooter(meta); + } + if (vectorData != null) { + CodecUtil.writeFooter(vectorData); + CodecUtil.writeFooter(vectorIndex); + } + } + + @Override + public long ramBytesUsed() { + long total = 0; + for (FieldWriter<?> field : fields) { + total += field.ramBytesUsed(); + } + return total; + } + + private void writeField(FieldWriter<?> fieldData, int maxDoc, long[] quantizedVecOffsetAndLen) + throws IOException { + // write vector values + long vectorDataOffset = vectorData.alignFilePointer(Float.BYTES); + switch (fieldData.fieldInfo.getVectorEncoding()) { + case BYTE -> writeByteVectors(fieldData); + case FLOAT32 -> writeFloat32Vectors(fieldData); + } + long vectorDataLength = vectorData.getFilePointer() - vectorDataOffset; + + // write graph + long vectorIndexOffset = vectorIndex.getFilePointer(); + OnHeapHnswGraph graph = fieldData.getGraph(); + int[][] graphLevelNodeOffsets = writeGraph(graph); + long vectorIndexLength = vectorIndex.getFilePointer() - vectorIndexOffset; + + writeMeta( + fieldData.isQuantized(), + fieldData.fieldInfo, + maxDoc, + fieldData.getConfiguredQuantile(), + fieldData.getMinQuantile(), + fieldData.getMaxQuantile(), + quantizedVecOffsetAndLen, + vectorDataOffset, + vectorDataLength, + vectorIndexOffset, + vectorIndexLength, + fieldData.docsWithField, + graph, + graphLevelNodeOffsets); + } + + private void writeFloat32Vectors(FieldWriter<?> fieldData) throws IOException { + final ByteBuffer buffer = + ByteBuffer.allocate(fieldData.dim * Float.BYTES).order(ByteOrder.LITTLE_ENDIAN); + for (Object v : fieldData.vectors) { + buffer.asFloatBuffer().put((float[]) v); + vectorData.writeBytes(buffer.array(), buffer.array().length); + } + } + + private void writeByteVectors(FieldWriter<?> fieldData) throws IOException { + for (Object v : fieldData.vectors) { + byte[] vector = (byte[]) v; + vectorData.writeBytes(vector, vector.length); + } + } + + private void writeSortingField( + FieldWriter<?> fieldData, + int maxDoc, + Sorter.DocMap sortMap, + long[] quantizedVectorOffsetAndLen) + throws IOException { + final int[] docIdOffsets = new int[sortMap.size()]; + int offset = 1; // 0 means no vector for this (field, document) + DocIdSetIterator iterator = fieldData.docsWithField.iterator(); + for (int docID = iterator.nextDoc(); + docID != DocIdSetIterator.NO_MORE_DOCS; + docID = iterator.nextDoc()) { + int newDocID = sortMap.oldToNew(docID); + docIdOffsets[newDocID] = offset++; + } + DocsWithFieldSet newDocsWithField = new DocsWithFieldSet(); + final int[] ordMap = new int[offset - 1]; // new ord to old ord + final int[] oldOrdMap = new int[offset - 1]; // old ord to new ord + int ord = 0; + int doc = 0; + for (int docIdOffset : docIdOffsets) { + if (docIdOffset != 0) { + ordMap[ord] = docIdOffset - 1; + oldOrdMap[docIdOffset - 1] = ord; + newDocsWithField.add(doc); + ord++; + } + doc++; + } + + // write vector values + long vectorDataOffset = + switch (fieldData.fieldInfo.getVectorEncoding()) { + case BYTE -> writeSortedByteVectors(fieldData, ordMap); + case FLOAT32 -> writeSortedFloat32Vectors(fieldData, ordMap); + }; + long vectorDataLength = vectorData.getFilePointer() - vectorDataOffset; + + // write graph + long vectorIndexOffset = vectorIndex.getFilePointer(); + OnHeapHnswGraph graph = fieldData.getGraph(); + int[][] graphLevelNodeOffsets = graph == null ? new int[0][] : new int[graph.numLevels()][]; + HnswGraph mockGraph = reconstructAndWriteGraph(graph, ordMap, oldOrdMap, graphLevelNodeOffsets); + long vectorIndexLength = vectorIndex.getFilePointer() - vectorIndexOffset; + + writeMeta( + fieldData.isQuantized(), + fieldData.fieldInfo, + maxDoc, + fieldData.getConfiguredQuantile(), + fieldData.getMinQuantile(), + fieldData.getMaxQuantile(), + quantizedVectorOffsetAndLen, + vectorDataOffset, + vectorDataLength, + vectorIndexOffset, + vectorIndexLength, + newDocsWithField, + mockGraph, + graphLevelNodeOffsets); + } + + private long writeSortedFloat32Vectors(FieldWriter<?> fieldData, int[] ordMap) + throws IOException { + long vectorDataOffset = vectorData.alignFilePointer(Float.BYTES); + final ByteBuffer buffer = + ByteBuffer.allocate(fieldData.dim * Float.BYTES).order(ByteOrder.LITTLE_ENDIAN); + for (int ordinal : ordMap) { + float[] vector = (float[]) fieldData.vectors.get(ordinal); + buffer.asFloatBuffer().put(vector); + vectorData.writeBytes(buffer.array(), buffer.array().length); + } + return vectorDataOffset; + } + + private long writeSortedByteVectors(FieldWriter<?> fieldData, int[] ordMap) throws IOException { + long vectorDataOffset = vectorData.alignFilePointer(Float.BYTES); + for (int ordinal : ordMap) { + byte[] vector = (byte[]) fieldData.vectors.get(ordinal); + vectorData.writeBytes(vector, vector.length); + } + return vectorDataOffset; + } + + /** + * Reconstructs the graph given the old and new node ids. + * + * <p>Additionally, the graph node connections are written to the vectorIndex. + * + * @param graph The current on heap graph + * @param newToOldMap the new node ids indexed to the old node ids + * @param oldToNewMap the old node ids indexed to the new node ids + * @param levelNodeOffsets where to place the new offsets for the nodes in the vector index. + * @return The graph + * @throws IOException if writing to vectorIndex fails + */ + private HnswGraph reconstructAndWriteGraph( + OnHeapHnswGraph graph, int[] newToOldMap, int[] oldToNewMap, int[][] levelNodeOffsets) + throws IOException { + if (graph == null) return null; + + List<int[]> nodesByLevel = new ArrayList<>(graph.numLevels()); + nodesByLevel.add(null); + + int maxOrd = graph.size(); + NodesIterator nodesOnLevel0 = graph.getNodesOnLevel(0); + levelNodeOffsets[0] = new int[nodesOnLevel0.size()]; + while (nodesOnLevel0.hasNext()) { + int node = nodesOnLevel0.nextInt(); + NeighborArray neighbors = graph.getNeighbors(0, newToOldMap[node]); + long offset = vectorIndex.getFilePointer(); + reconstructAndWriteNeigbours(neighbors, oldToNewMap, maxOrd); + levelNodeOffsets[0][node] = Math.toIntExact(vectorIndex.getFilePointer() - offset); + } + + for (int level = 1; level < graph.numLevels(); level++) { + NodesIterator nodesOnLevel = graph.getNodesOnLevel(level); + int[] newNodes = new int[nodesOnLevel.size()]; + for (int n = 0; nodesOnLevel.hasNext(); n++) { + newNodes[n] = oldToNewMap[nodesOnLevel.nextInt()]; + } + Arrays.sort(newNodes); + nodesByLevel.add(newNodes); + levelNodeOffsets[level] = new int[newNodes.length]; + int nodeOffsetIndex = 0; + for (int node : newNodes) { + NeighborArray neighbors = graph.getNeighbors(level, newToOldMap[node]); + long offset = vectorIndex.getFilePointer(); + reconstructAndWriteNeigbours(neighbors, oldToNewMap, maxOrd); + levelNodeOffsets[level][nodeOffsetIndex++] = + Math.toIntExact(vectorIndex.getFilePointer() - offset); + } + } + return new HnswGraph() { + @Override + public int nextNeighbor() { + throw new UnsupportedOperationException("Not supported on a mock graph"); + } + + @Override + public void seek(int level, int target) { + throw new UnsupportedOperationException("Not supported on a mock graph"); + } + + @Override + public int size() { + return graph.size(); + } + + @Override + public int numLevels() { + return graph.numLevels(); + } + + @Override + public int entryNode() { + throw new UnsupportedOperationException("Not supported on a mock graph"); + } + + @Override + public NodesIterator getNodesOnLevel(int level) { + if (level == 0) { + return graph.getNodesOnLevel(0); + } else { + return new ArrayNodesIterator(nodesByLevel.get(level), nodesByLevel.get(level).length); + } + } + }; + } + + private void reconstructAndWriteNeigbours(NeighborArray neighbors, int[] oldToNewMap, int maxOrd) + throws IOException { + int size = neighbors.size(); + vectorIndex.writeVInt(size); + + // Destructively modify; it's ok we are discarding it after this + int[] nnodes = neighbors.node(); + for (int i = 0; i < size; i++) { + nnodes[i] = oldToNewMap[nnodes[i]]; + } + Arrays.sort(nnodes, 0, size); + // Now that we have sorted, do delta encoding to minimize the required bits to store the + // information + for (int i = size - 1; i > 0; --i) { + assert nnodes[i] < maxOrd : "node too large: " + nnodes[i] + ">=" + maxOrd; + nnodes[i] -= nnodes[i - 1]; + } + for (int i = 0; i < size; i++) { + vectorIndex.writeVInt(nnodes[i]); + } + } + + @Override + public void mergeOneField(FieldInfo fieldInfo, MergeState mergeState) throws IOException { + long vectorDataOffset = vectorData.alignFilePointer(Float.BYTES); + IndexOutput tempVectorData = + segmentWriteState.directory.createTempOutput( + vectorData.getName(), "temp", segmentWriteState.context); + IndexInput vectorDataInput = null; + IndexInput quantizationDataInput = null; + Lucene99ScalarQuantizedVectorsWriter.MergedQuantileState quantizationState = null; + boolean success = false; + try { + long[] quantizedVectorDataOffsetAndLength = null; + if (quantizedVectorsWriter != null) { + quantizationState = + quantizedVectorsWriter.mergeOneField(segmentWriteState, fieldInfo, mergeState); + if (quantizationState != null) { + quantizedVectorDataOffsetAndLength = new long[2]; + quantizedVectorDataOffsetAndLength[0] = quantizedVectorData.alignFilePointer(Float.BYTES); + quantizationDataInput = + segmentWriteState.directory.openInput( + quantizationState.tempVectorFileName, segmentWriteState.context); + quantizedVectorData.copyBytes( + quantizationDataInput, quantizationDataInput.length() - CodecUtil.footerLength()); + quantizedVectorDataOffsetAndLength[1] = + quantizedVectorData.getFilePointer() - quantizedVectorDataOffsetAndLength[0]; + CodecUtil.retrieveChecksum(quantizationDataInput); + } + } + // write the vector data to a temporary file + DocsWithFieldSet docsWithField = + switch (fieldInfo.getVectorEncoding()) { + case BYTE -> writeByteVectorData( + tempVectorData, MergedVectorValues.mergeByteVectorValues(fieldInfo, mergeState)); + case FLOAT32 -> writeVectorData( + tempVectorData, MergedVectorValues.mergeFloatVectorValues(fieldInfo, mergeState)); + }; + CodecUtil.writeFooter(tempVectorData); + IOUtils.close(tempVectorData); + + // copy the temporary file vectors to the actual data file + vectorDataInput = + segmentWriteState.directory.openInput( + tempVectorData.getName(), segmentWriteState.context); + vectorData.copyBytes(vectorDataInput, vectorDataInput.length() - CodecUtil.footerLength()); + CodecUtil.retrieveChecksum(vectorDataInput); + long vectorDataLength = vectorData.getFilePointer() - vectorDataOffset; + long vectorIndexOffset = vectorIndex.getFilePointer(); + // build the graph using the temporary vector data + // we use Lucene99HnswVectorsReader.DenseOffHeapVectorValues for the graph construction + // doesn't need to know docIds + // TODO: separate random access vector values from DocIdSetIterator? + int byteSize = fieldInfo.getVectorDimension() * fieldInfo.getVectorEncoding().byteSize; + OnHeapHnswGraph graph = null; + int[][] vectorIndexNodeOffsets = null; + if (docsWithField.cardinality() != 0) { + // build graph + int initializerIndex = selectGraphForInitialization(mergeState, fieldInfo); + graph = + switch (fieldInfo.getVectorEncoding()) { + case BYTE -> { + OffHeapByteVectorValues.DenseOffHeapVectorValues vectorValues = + new OffHeapByteVectorValues.DenseOffHeapVectorValues( + fieldInfo.getVectorDimension(), + docsWithField.cardinality(), + vectorDataInput, + byteSize); + RandomVectorScorerSupplier scorerSupplier = + RandomVectorScorerSupplier.createBytes( + vectorValues, fieldInfo.getVectorSimilarityFunction()); + HnswGraphBuilder hnswGraphBuilder = + createHnswGraphBuilder(mergeState, fieldInfo, scorerSupplier, initializerIndex); + hnswGraphBuilder.setInfoStream(segmentWriteState.infoStream); + yield hnswGraphBuilder.build(vectorValues.size()); + } + case FLOAT32 -> { + final int maxOrd; + final RandomVectorScorerSupplier scorerSupplier; + if (quantizationState != null) { + RandomAccessQuantizedByteVectorValues values = + new OffHeapQuantizedByteVectorValues.DenseOffHeapVectorValues( + fieldInfo.getVectorDimension(), + docsWithField.cardinality(), + quantizationDataInput); + maxOrd = values.size(); + scorerSupplier = + new ScalarQuantizedRandomVectorScorerSupplier( + fieldInfo.getVectorSimilarityFunction(), + quantizationState.mergeQuantile, + values); + } else { + OffHeapFloatVectorValues.DenseOffHeapVectorValues vectorValues = + new OffHeapFloatVectorValues.DenseOffHeapVectorValues( + fieldInfo.getVectorDimension(), + docsWithField.cardinality(), + vectorDataInput, + byteSize); + maxOrd = vectorValues.size(); + scorerSupplier = + RandomVectorScorerSupplier.createFloats( + vectorValues, fieldInfo.getVectorSimilarityFunction()); + } + HnswGraphBuilder hnswGraphBuilder = + createHnswGraphBuilder(mergeState, fieldInfo, scorerSupplier, initializerIndex); + hnswGraphBuilder.setInfoStream(segmentWriteState.infoStream); + yield hnswGraphBuilder.build(maxOrd); + } + }; + vectorIndexNodeOffsets = writeGraph(graph); + } + long vectorIndexLength = vectorIndex.getFilePointer() - vectorIndexOffset; + writeMeta( + quantizationState != null, + fieldInfo, + segmentWriteState.segmentInfo.maxDoc(), + quantizationState == null ? null : quantizationState.configuredQuantile, + quantizationState == null ? null : quantizationState.mergeQuantile.getLowerQuantile(), + quantizationState == null ? null : quantizationState.mergeQuantile.getUpperQuantile(), + quantizedVectorDataOffsetAndLength, + vectorDataOffset, + vectorDataLength, + vectorIndexOffset, + vectorIndexLength, + docsWithField, + graph, + vectorIndexNodeOffsets); + success = true; + } finally { + IOUtils.close(vectorDataInput, quantizationDataInput); + if (success) { + segmentWriteState.directory.deleteFile(tempVectorData.getName()); + if (quantizationState != null) { + segmentWriteState.directory.deleteFile(quantizationState.tempVectorFileName); + } + } else { + IOUtils.closeWhileHandlingException(tempVectorData); + IOUtils.deleteFilesIgnoringExceptions( + segmentWriteState.directory, tempVectorData.getName()); + } + } + } + + private HnswGraphBuilder createHnswGraphBuilder( + MergeState mergeState, + FieldInfo fieldInfo, + RandomVectorScorerSupplier scorerSupplier, + int initializerIndex) + throws IOException { + if (initializerIndex == -1) { + return HnswGraphBuilder.create(scorerSupplier, M, beamWidth, HnswGraphBuilder.randSeed); + } + + HnswGraph initializerGraph = + getHnswGraphFromReader(fieldInfo.name, mergeState.knnVectorsReaders[initializerIndex]); + Map<Integer, Integer> ordinalMapper = + getOldToNewOrdinalMap(mergeState, fieldInfo, initializerIndex); + return HnswGraphBuilder.create( + scorerSupplier, M, beamWidth, HnswGraphBuilder.randSeed, initializerGraph, ordinalMapper); + } + + private int selectGraphForInitialization(MergeState mergeState, FieldInfo fieldInfo) + throws IOException { + // Find the KnnVectorReader with the most docs that meets the following criteria: + // 1. Does not contain any deleted docs + // 2. Is a Lucene99HnswVectorsReader/PerFieldKnnVectorReader + // If no readers exist that meet this criteria, return -1. If they do, return their index in + // merge state + int maxCandidateVectorCount = 0; + int initializerIndex = -1; + + for (int i = 0; i < mergeState.liveDocs.length; i++) { + KnnVectorsReader currKnnVectorsReader = mergeState.knnVectorsReaders[i]; + if (mergeState.knnVectorsReaders[i] + instanceof PerFieldKnnVectorsFormat.FieldsReader candidateReader) { + currKnnVectorsReader = candidateReader.getFieldReader(fieldInfo.name); + } + + if (!allMatch(mergeState.liveDocs[i]) + || !(currKnnVectorsReader instanceof HnswGraphProvider)) { + continue; + } + + int candidateVectorCount = 0; + switch (fieldInfo.getVectorEncoding()) { + case BYTE -> { + ByteVectorValues byteVectorValues = + currKnnVectorsReader.getByteVectorValues(fieldInfo.name); + if (byteVectorValues == null) { + continue; + } + candidateVectorCount = byteVectorValues.size(); + } + case FLOAT32 -> { + FloatVectorValues vectorValues = + currKnnVectorsReader.getFloatVectorValues(fieldInfo.name); + if (vectorValues == null) { + continue; + } + candidateVectorCount = vectorValues.size(); + } + } + + if (candidateVectorCount > maxCandidateVectorCount) { + maxCandidateVectorCount = candidateVectorCount; + initializerIndex = i; + } + } + return initializerIndex; + } + + private HnswGraph getHnswGraphFromReader(String fieldName, KnnVectorsReader knnVectorsReader) + throws IOException { + if (knnVectorsReader instanceof PerFieldKnnVectorsFormat.FieldsReader perFieldReader + && perFieldReader.getFieldReader(fieldName) instanceof HnswGraphProvider fieldReader) { + return fieldReader.getGraph(fieldName); + } + + if (knnVectorsReader instanceof HnswGraphProvider) { + return ((HnswGraphProvider) knnVectorsReader).getGraph(fieldName); + } + + // We should not reach here because knnVectorsReader's type is checked in + // selectGraphForInitialization + throw new IllegalArgumentException( + "Invalid KnnVectorsReader type for field: " + + fieldName + + ". Must be Lucene99HnswVectorsReader or newer"); + } + + private Map<Integer, Integer> getOldToNewOrdinalMap( + MergeState mergeState, FieldInfo fieldInfo, int initializerIndex) throws IOException { + + DocIdSetIterator initializerIterator = null; + + switch (fieldInfo.getVectorEncoding()) { + case BYTE -> initializerIterator = + mergeState.knnVectorsReaders[initializerIndex].getByteVectorValues(fieldInfo.name); + case FLOAT32 -> initializerIterator = + mergeState.knnVectorsReaders[initializerIndex].getFloatVectorValues(fieldInfo.name); + } + + MergeState.DocMap initializerDocMap = mergeState.docMaps[initializerIndex]; + + Map<Integer, Integer> newIdToOldOrdinal = new HashMap<>(); + int oldOrd = 0; + int maxNewDocID = -1; + for (int oldId = initializerIterator.nextDoc(); + oldId != NO_MORE_DOCS; + oldId = initializerIterator.nextDoc()) { + if (isCurrentVectorNull(initializerIterator)) { + continue; + } + int newId = initializerDocMap.get(oldId); + maxNewDocID = Math.max(newId, maxNewDocID); + newIdToOldOrdinal.put(newId, oldOrd); + oldOrd++; + } + + if (maxNewDocID == -1) { + return Collections.emptyMap(); + } + + Map<Integer, Integer> oldToNewOrdinalMap = new HashMap<>(); + + DocIdSetIterator vectorIterator = null; + switch (fieldInfo.getVectorEncoding()) { + case BYTE -> vectorIterator = MergedVectorValues.mergeByteVectorValues(fieldInfo, mergeState); + case FLOAT32 -> vectorIterator = + MergedVectorValues.mergeFloatVectorValues(fieldInfo, mergeState); + } + + int newOrd = 0; + for (int newDocId = vectorIterator.nextDoc(); + newDocId <= maxNewDocID; + newDocId = vectorIterator.nextDoc()) { + if (isCurrentVectorNull(vectorIterator)) { + continue; + } + + if (newIdToOldOrdinal.containsKey(newDocId)) { + oldToNewOrdinalMap.put(newIdToOldOrdinal.get(newDocId), newOrd); + } + newOrd++; + } + + return oldToNewOrdinalMap; + } + + private boolean isCurrentVectorNull(DocIdSetIterator docIdSetIterator) throws IOException { + if (docIdSetIterator instanceof FloatVectorValues) { + return ((FloatVectorValues) docIdSetIterator).vectorValue() == null; + } + + if (docIdSetIterator instanceof ByteVectorValues) { + return ((ByteVectorValues) docIdSetIterator).vectorValue() == null; + } + + return true; + } + + private boolean allMatch(Bits bits) { + if (bits == null) { + return true; + } + + for (int i = 0; i < bits.length(); i++) { + if (!bits.get(i)) { + return false; + } + } + return true; + } + + /** + * @param graph Write the graph in a compressed format + * @return The non-cumulative offsets for the nodes. Should be used to create cumulative offsets. + * @throws IOException if writing to vectorIndex fails + */ + private int[][] writeGraph(OnHeapHnswGraph graph) throws IOException { + if (graph == null) return new int[0][0]; + // write vectors' neighbours on each level into the vectorIndex file + int countOnLevel0 = graph.size(); + int[][] offsets = new int[graph.numLevels()][]; + for (int level = 0; level < graph.numLevels(); level++) { + int[] sortedNodes = getSortedNodes(graph.getNodesOnLevel(level)); + offsets[level] = new int[sortedNodes.length]; + int nodeOffsetId = 0; + for (int node : sortedNodes) { + NeighborArray neighbors = graph.getNeighbors(level, node); + int size = neighbors.size(); + // Write size in VInt as the neighbors list is typically small + long offsetStart = vectorIndex.getFilePointer(); + vectorIndex.writeVInt(size); + // Destructively modify; it's ok we are discarding it after this + int[] nnodes = neighbors.node(); + Arrays.sort(nnodes, 0, size); + // Now that we have sorted, do delta encoding to minimize the required bits to store the + // information + for (int i = size - 1; i > 0; --i) { + assert nnodes[i] < countOnLevel0 : "node too large: " + nnodes[i] + ">=" + countOnLevel0; + nnodes[i] -= nnodes[i - 1]; + } + for (int i = 0; i < size; i++) { + vectorIndex.writeVInt(nnodes[i]); + } + offsets[level][nodeOffsetId++] = + Math.toIntExact(vectorIndex.getFilePointer() - offsetStart); + } + } + return offsets; + } + + public static int[] getSortedNodes(NodesIterator nodesOnLevel) { + int[] sortedNodes = new int[nodesOnLevel.size()]; + for (int n = 0; nodesOnLevel.hasNext(); n++) { + sortedNodes[n] = nodesOnLevel.nextInt(); + } + Arrays.sort(sortedNodes); + return sortedNodes; + } + + private void writeMeta( + boolean isQuantized, + FieldInfo field, + int maxDoc, + Float configuredQuantizationQuantile, + Float lowerQuantile, + Float upperQuantile, + long[] quantizedVectorDataOffsetAndLen, + long vectorDataOffset, + long vectorDataLength, + long vectorIndexOffset, + long vectorIndexLength, + DocsWithFieldSet docsWithField, + HnswGraph graph, + int[][] graphLevelNodeOffsets) + throws IOException { + meta.writeInt(field.number); + meta.writeInt(field.getVectorEncoding().ordinal()); + meta.writeInt(field.getVectorSimilarityFunction().ordinal()); + meta.writeByte(isQuantized ? (byte) 1 : (byte) 0); + if (isQuantized) { + assert lowerQuantile != null + && upperQuantile != null + && quantizedVectorDataOffsetAndLen != null; + assert quantizedVectorDataOffsetAndLen.length == 2; + meta.writeInt( + Float.floatToIntBits( + configuredQuantizationQuantile != null + ? configuredQuantizationQuantile + : calculateDefaultQuantile(field.getVectorDimension()))); + meta.writeInt(Float.floatToIntBits(lowerQuantile)); + meta.writeInt(Float.floatToIntBits(upperQuantile)); + meta.writeVLong(quantizedVectorDataOffsetAndLen[0]); + meta.writeVLong(quantizedVectorDataOffsetAndLen[1]); + } else { + assert configuredQuantizationQuantile == null + && lowerQuantile == null + && upperQuantile == null + && quantizedVectorDataOffsetAndLen == null; + } + meta.writeVLong(vectorDataOffset); + meta.writeVLong(vectorDataLength); + meta.writeVLong(vectorIndexOffset); + meta.writeVLong(vectorIndexLength); + meta.writeVInt(field.getVectorDimension()); + + // write docIDs + int count = docsWithField.cardinality(); + meta.writeInt(count); + if (isQuantized) { + writeOrdToDoc(quantizedVectorData, docsWithField, count, maxDoc); Review Comment: > The raw vectors file is already opened so not sure I follow. I mean, not have to seek in the file. I would have thought seeking the file would page it into cache. -- 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-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
