Hierarchical centroid storage for DiskBBQ (#132010)

This commit presents a hierarchical layer on top of the DiskBBQ centroids to reduce the number of centroids scored
at search time.

Author: iverase

qa/vector/src/main/java/org/elasticsearch/test/knn/KnnIndexTester.java

@@ -101,7 +101,7 @@ private static String formatIndexPath(CmdLineArgs args) {
     static Codec createCodec(CmdLineArgs args) {
         final KnnVectorsFormat format;
         if (args.indexType() == IndexType.IVF) {
-            format = new IVFVectorsFormat(args.ivfClusterSize());
+            format = new IVFVectorsFormat(args.ivfClusterSize(), IVFVectorsFormat.DEFAULT_CENTROIDS_PER_PARENT_CLUSTER);
         } else {
             if (args.quantizeBits() == 1) {
                 if (args.indexType() == IndexType.FLAT) {

server/src/main/java/org/elasticsearch/index/codec/vectors/DefaultIVFVectorsReader.java

@@ -29,8 +29,6 @@
 
 import static org.apache.lucene.codecs.lucene102.Lucene102BinaryQuantizedVectorsFormat.QUERY_BITS;
 import static org.apache.lucene.index.VectorSimilarityFunction.COSINE;
-import static org.apache.lucene.index.VectorSimilarityFunction.EUCLIDEAN;
-import static org.apache.lucene.index.VectorSimilarityFunction.MAXIMUM_INNER_PRODUCT;
 import static org.elasticsearch.index.codec.vectors.BQSpaceUtils.transposeHalfByte;
 import static org.elasticsearch.index.codec.vectors.BQVectorUtils.discretize;
 import static org.elasticsearch.index.codec.vectors.OptimizedScalarQuantizer.DEFAULT_LAMBDA;
@@ -41,7 +39,9 @@
  * brute force and then scores the top ones using the posting list.
  */
 public class DefaultIVFVectorsReader extends IVFVectorsReader implements OffHeapStats {
-    private static final float FOUR_BIT_SCALE = 1f / ((1 << 4) - 1);
+
+    // The percentage of centroids that are scored to keep recall
+    public static final double CENTROID_SAMPLING_PERCENTAGE = 0.2;
 
     public DefaultIVFVectorsReader(SegmentReadState state, FlatVectorsReader rawVectorsReader) throws IOException {
         super(state, rawVectorsReader);
@@ -54,8 +54,12 @@ CentroidIterator getCentroidIterator(FieldInfo fieldInfo, int numCentroids, Inde
         final float globalCentroidDp = fieldEntry.globalCentroidDp();
         final OptimizedScalarQuantizer scalarQuantizer = new OptimizedScalarQuantizer(fieldInfo.getVectorSimilarityFunction());
         final int[] scratch = new int[targetQuery.length];
+        float[] targetQueryCopy = ArrayUtil.copyArray(targetQuery);
+        if (fieldInfo.getVectorSimilarityFunction() == COSINE) {
+            VectorUtil.l2normalize(targetQueryCopy);
+        }
         final OptimizedScalarQuantizer.QuantizationResult queryParams = scalarQuantizer.scalarQuantize(
-            ArrayUtil.copyArray(targetQuery),
+            targetQueryCopy,
             scratch,
             (byte) 4,
             fieldEntry.globalCentroid()
@@ -65,67 +69,227 @@ CentroidIterator getCentroidIterator(FieldInfo fieldInfo, int numCentroids, Inde
             quantized[i] = (byte) scratch[i];
         }
         final ES91Int4VectorsScorer scorer = ESVectorUtil.getES91Int4VectorsScorer(centroids, fieldInfo.getVectorDimension());
-        NeighborQueue queue = new NeighborQueue(fieldEntry.numCentroids(), true);
         centroids.seek(0L);
-        final float[] centroidCorrectiveValues = new float[3];
-        for (int i = 0; i < numCentroids; i++) {
-            final float qcDist = scorer.int4DotProduct(quantized);
-            centroids.readFloats(centroidCorrectiveValues, 0, 3);
-            final int quantizedCentroidComponentSum = Short.toUnsignedInt(centroids.readShort());
-            float score = int4QuantizedScore(
-                qcDist,
+        int numParents = centroids.readVInt();
+        if (numParents > 0) {
+            return getCentroidIteratorWithParents(
+                fieldInfo,
+                centroids,
+                numParents,
+                numCentroids,
+                scorer,
+                quantized,
                 queryParams,
-                fieldInfo.getVectorDimension(),
-                centroidCorrectiveValues,
-                quantizedCentroidComponentSum,
-                globalCentroidDp,
-                fieldInfo.getVectorSimilarityFunction()
+                globalCentroidDp
             );
-            queue.add(i, score);
         }
-        final long offset = centroids.getFilePointer();
+        return getCentroidIteratorNoParent(fieldInfo, centroids, numCentroids, scorer, quantized, queryParams, globalCentroidDp);
+    }
+
+    private static CentroidIterator getCentroidIteratorNoParent(
+        FieldInfo fieldInfo,
+        IndexInput centroids,
+        int numCentroids,
+        ES91Int4VectorsScorer scorer,
+        byte[] quantizeQuery,
+        OptimizedScalarQuantizer.QuantizationResult queryParams,
+        float globalCentroidDp
+    ) throws IOException {
+        final NeighborQueue neighborQueue = new NeighborQueue(numCentroids, true);
+        score(
+            neighborQueue,
+            numCentroids,
+            0,
+            scorer,
+            quantizeQuery,
+            queryParams,
+            globalCentroidDp,
+            fieldInfo.getVectorSimilarityFunction(),
+            new float[ES91Int4VectorsScorer.BULK_SIZE]
+        );
+        long offset = centroids.getFilePointer();
         return new CentroidIterator() {
             @Override
             public boolean hasNext() {
-                return queue.size() > 0;
+                return neighborQueue.size() > 0;
             }
 
             @Override
             public long nextPostingListOffset() throws IOException {
-                int centroidOrdinal = queue.pop();
+                int centroidOrdinal = neighborQueue.pop();
                 centroids.seek(offset + (long) Long.BYTES * centroidOrdinal);
                 return centroids.readLong();
             }
         };
     }
 
-    // TODO can we do this in off-heap blocks?
-    private float int4QuantizedScore(
-        float qcDist,
+    private static CentroidIterator getCentroidIteratorWithParents(
+        FieldInfo fieldInfo,
+        IndexInput centroids,
+        int numParents,
+        int numCentroids,
+        ES91Int4VectorsScorer scorer,
+        byte[] quantizeQuery,
+        OptimizedScalarQuantizer.QuantizationResult queryParams,
+        float globalCentroidDp
+    ) throws IOException {
+        // build the three queues we are going to use
+        final NeighborQueue parentsQueue = new NeighborQueue(numParents, true);
+        final int maxChildrenSize = centroids.readVInt();
+        final NeighborQueue currentParentQueue = new NeighborQueue(maxChildrenSize, true);
+        final int bufferSize = (int) Math.max(numCentroids * CENTROID_SAMPLING_PERCENTAGE, 1);
+        final NeighborQueue neighborQueue = new NeighborQueue(bufferSize, true);
+        // score the parents
+        final float[] scores = new float[ES91Int4VectorsScorer.BULK_SIZE];
+        score(
+            parentsQueue,
+            numParents,
+            0,
+            scorer,
+            quantizeQuery,
+            queryParams,
+            globalCentroidDp,
+            fieldInfo.getVectorSimilarityFunction(),
+            scores
+        );
+        final long centroidQuantizeSize = fieldInfo.getVectorDimension() + 3 * Float.BYTES + Short.BYTES;
+        final long offset = centroids.getFilePointer();
+        final long childrenOffset = offset + (long) Long.BYTES * numParents;
+        // populate the children's queue by reading parents one by one
+        while (parentsQueue.size() > 0 && neighborQueue.size() < bufferSize) {
+            final int pop = parentsQueue.pop();
+            populateOneChildrenGroup(
+                currentParentQueue,
+                centroids,
+                offset + 2L * Integer.BYTES * pop,
+                childrenOffset,
+                centroidQuantizeSize,
+                fieldInfo,
+                scorer,
+                quantizeQuery,
+                queryParams,
+                globalCentroidDp,
+                scores
+            );
+            while (currentParentQueue.size() > 0 && neighborQueue.size() < bufferSize) {
+                final float score = currentParentQueue.topScore();
+                final int children = currentParentQueue.pop();
+                neighborQueue.add(children, score);
+            }
+        }
+        final long childrenFileOffsets = childrenOffset + centroidQuantizeSize * numCentroids;
+        return new CentroidIterator() {
+            @Override
+            public boolean hasNext() {
+                return neighborQueue.size() > 0;
+            }
+
+            @Override
+            public long nextPostingListOffset() throws IOException {
+                int centroidOrdinal = neighborQueue.pop();
+                updateQueue(); // add one children if available so the queue remains fully populated
+                centroids.seek(childrenFileOffsets + (long) Long.BYTES * centroidOrdinal);
+                return centroids.readLong();
+            }
+
+            private void updateQueue() throws IOException {
+                if (currentParentQueue.size() > 0) {
+                    // add a children from the current parent queue
+                    float score = currentParentQueue.topScore();
+                    int children = currentParentQueue.pop();
+                    neighborQueue.add(children, score);
+                } else if (parentsQueue.size() > 0) {
+                    // add a new parent from the current parent queue
+                    int pop = parentsQueue.pop();
+                    populateOneChildrenGroup(
+                        currentParentQueue,
+                        centroids,
+                        offset + 2L * Integer.BYTES * pop,
+                        childrenOffset,
+                        centroidQuantizeSize,
+                        fieldInfo,
+                        scorer,
+                        quantizeQuery,
+                        queryParams,
+                        globalCentroidDp,
+                        scores
+                    );
+                    updateQueue();
+                }
+            }
+        };
+    }
+
+    private static void populateOneChildrenGroup(
+        NeighborQueue neighborQueue,
+        IndexInput centroids,
+        long parentOffset,
+        long childrenOffset,
+        long centroidQuantizeSize,
+        FieldInfo fieldInfo,
+        ES91Int4VectorsScorer scorer,
+        byte[] quantizeQuery,
+        OptimizedScalarQuantizer.QuantizationResult queryParams,
+        float globalCentroidDp,
+        float[] scores
+    ) throws IOException {
+        centroids.seek(parentOffset);
+        int childrenOrdinal = centroids.readInt();
+        int numChildren = centroids.readInt();
+        centroids.seek(childrenOffset + centroidQuantizeSize * childrenOrdinal);
+        score(
+            neighborQueue,
+            numChildren,
+            childrenOrdinal,
+            scorer,
+            quantizeQuery,
+            queryParams,
+            globalCentroidDp,
+            fieldInfo.getVectorSimilarityFunction(),
+            scores
+        );
+    }
+
+    private static void score(
+        NeighborQueue neighborQueue,
+        int size,
+        int scoresOffset,
+        ES91Int4VectorsScorer scorer,
+        byte[] quantizeQuery,
         OptimizedScalarQuantizer.QuantizationResult queryCorrections,
-        int dims,
-        float[] targetCorrections,
-        int targetComponentSum,
         float centroidDp,
-        VectorSimilarityFunction similarityFunction
-    ) {
-        float ax = targetCorrections[0];
-        float lx = (targetCorrections[1] - ax) * FOUR_BIT_SCALE;
-        float ay = queryCorrections.lowerInterval();
-        float ly = (queryCorrections.upperInterval() - ay) * FOUR_BIT_SCALE;
-        float y1 = queryCorrections.quantizedComponentSum();
-        float score = ax * ay * dims + ay * lx * (float) targetComponentSum + ax * ly * y1 + lx * ly * qcDist;
-        if (similarityFunction == EUCLIDEAN) {
-            score = queryCorrections.additionalCorrection() + targetCorrections[2] - 2 * score;
-            return Math.max(1 / (1f + score), 0);
-        } else {
-            // For cosine and max inner product, we need to apply the additional correction, which is
-            // assumed to be the non-centered dot-product between the vector and the centroid
-            score += queryCorrections.additionalCorrection() + targetCorrections[2] - centroidDp;
-            if (similarityFunction == MAXIMUM_INNER_PRODUCT) {
-                return VectorUtil.scaleMaxInnerProductScore(score);
+        VectorSimilarityFunction similarityFunction,
+        float[] scores
+    ) throws IOException {
+        int limit = size - ES91Int4VectorsScorer.BULK_SIZE + 1;
+        int i = 0;
+        for (; i < limit; i += ES91Int4VectorsScorer.BULK_SIZE) {
+            scorer.scoreBulk(
+                quantizeQuery,
+                queryCorrections.lowerInterval(),
+                queryCorrections.upperInterval(),
+                queryCorrections.quantizedComponentSum(),
+                queryCorrections.additionalCorrection(),
+                similarityFunction,
+                centroidDp,
+                scores
+            );
+            for (int j = 0; j < ES91Int4VectorsScorer.BULK_SIZE; j++) {
+                neighborQueue.add(scoresOffset + i + j, scores[j]);
             }
-            return Math.max((1f + score) / 2f, 0);
+        }
+
+        for (; i < size; i++) {
+            float score = scorer.score(
+                quantizeQuery,
+                queryCorrections.lowerInterval(),
+                queryCorrections.upperInterval(),
+                queryCorrections.quantizedComponentSum(),
+                queryCorrections.additionalCorrection(),
+                similarityFunction,
+                centroidDp
+            );
+            neighborQueue.add(scoresOffset + i, score);
         }
     }