Skip to content

Diskbbq posting list order #132697

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
elasticsearchmachine merged 9 commits into from
Aug 12, 2025
Merged

Diskbbq posting list order #132697

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
4d2841e
Sort postings lists by doc ID
benwtrent Aug 11, 2025
3d4e7a0
Merge remote-tracking branch 'upstream/main' into diskbbq-posting-lis…
benwtrent Aug 11, 2025
8d8ea36
iter
benwtrent Aug 11, 2025
ba91845
delta encode
benwtrent Aug 11, 2025
757c149
Update DocIdsWriter.java
benwtrent Aug 12, 2025
61776d7
Merge branch 'main' into diskbbq-posting-list-order
benwtrent Aug 12, 2025
78643cd
[CI] Auto commit changes from spotless
Aug 12, 2025
7fa53d1
Update DefaultIVFVectorsWriter.java
benwtrent Aug 12, 2025
7137555
Merge remote-tracking branch 'upstream/main' into diskbbq-posting-lis…
benwtrent Aug 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions server/src/main/java/org/elasticsearch/index/codec/vectors/DefaultIVFVectorsReader.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,12 +326,12 @@ private static class MemorySegmentPostingsVisitor implements PostingVisitor {
final float[] centroid;
long slicePos;
OptimizedScalarQuantizer.QuantizationResult queryCorrections;
DocIdsWriter docIdsWriter = new DocIdsWriter();

final float[] scratch;
final int[] quantizationScratch;
final byte[] quantizedQueryScratch;
final OptimizedScalarQuantizer quantizer;
final DocIdsWriter idsWriter = new DocIdsWriter();
final float[] correctiveValues = new float[3];
final long quantizedVectorByteSize;

Expand Down Expand Up @@ -369,7 +369,13 @@ public int resetPostingsScorer(long offset) throws IOException {
vectors = indexInput.readVInt();
// read the doc ids
assert vectors <= docIdsScratch.length;
docIdsWriter.readInts(indexInput, vectors, docIdsScratch);
idsWriter.readInts(indexInput, vectors, docIdsScratch);
// reconstitute from the deltas
int sum = 0;
for (int i = 0; i < vectors; i++) {
sum += docIdsScratch[i];
docIdsScratch[i] = sum;
}
slicePos = indexInput.getFilePointer();
return vectors;
}
Expand Down
88 changes: 82 additions & 6 deletions server/src/main/java/org/elasticsearch/index/codec/vectors/DefaultIVFVectorsWriter.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
import org.apache.lucene.store.IOContext;
import org.apache.lucene.store.IndexInput;
import org.apache.lucene.store.IndexOutput;
import org.apache.lucene.util.IntroSorter;
import org.apache.lucene.util.LongValues;
import org.apache.lucene.util.VectorUtil;
import org.apache.lucene.util.hnsw.IntToIntFunction;
Expand Down Expand Up @@ -101,14 +102,17 @@ LongValues buildAndWritePostingsLists(
postingsOutput.writeVInt(maxPostingListSize);
// write the posting lists
final PackedLongValues.Builder offsets = PackedLongValues.monotonicBuilder(PackedInts.COMPACT);
DocIdsWriter docIdsWriter = new DocIdsWriter();
DiskBBQBulkWriter bulkWriter = new DiskBBQBulkWriter.OneBitDiskBBQBulkWriter(ES91OSQVectorsScorer.BULK_SIZE, postingsOutput);
OnHeapQuantizedVectors onHeapQuantizedVectors = new OnHeapQuantizedVectors(
floatVectorValues,
fieldInfo.getVectorDimension(),
new OptimizedScalarQuantizer(fieldInfo.getVectorSimilarityFunction())
);
final ByteBuffer buffer = ByteBuffer.allocate(fieldInfo.getVectorDimension() * Float.BYTES).order(ByteOrder.LITTLE_ENDIAN);
int[] docIds = null;
int[] docDeltas = null;
int[] clusterOrds = null;
DocIdsWriter idsWriter = new DocIdsWriter();
for (int c = 0; c < centroidSupplier.size(); c++) {
float[] centroid = centroidSupplier.centroid(c);
int[] cluster = assignmentsByCluster[c];
Expand All @@ -121,11 +125,29 @@ LongValues buildAndWritePostingsLists(
int size = cluster.length;
// write docIds
postingsOutput.writeVInt(size);
onHeapQuantizedVectors.reset(centroid, size, ord -> cluster[ord]);
if (docIds == null || docIds.length < cluster.length) {
docIds = new int[cluster.length];
clusterOrds = new int[cluster.length];
docDeltas = new int[cluster.length];
}
for (int j = 0; j < size; j++) {
docIds[j] = floatVectorValues.ordToDoc(cluster[j]);
clusterOrds[j] = j;
}
final int[] finalDocs = docIds;
final int[] finalOrds = clusterOrds;
// sort cluster.buffer by docIds values, this way cluster ordinals are sorted by docIds
new IntSorter(clusterOrds, i -> finalDocs[i]).sort(0, size);
// encode doc deltas
for (int j = 0; j < size; j++) {
docDeltas[j] = j == 0 ? finalDocs[finalOrds[j]] : finalDocs[finalOrds[j]] - finalDocs[finalOrds[j - 1]];
}
final int[] finalDocDeltas = docDeltas;
onHeapQuantizedVectors.reset(centroid, size, ord -> cluster[finalOrds[ord]]);
// TODO we might want to consider putting the docIds in a separate file
// to aid with only having to fetch vectors from slower storage when they are required
// keeping them in the same file indicates we pull the entire file into cache
docIdsWriter.writeDocIds(j -> floatVectorValues.ordToDoc(cluster[j]), size, postingsOutput);
idsWriter.writeDocIds(i -> finalDocDeltas[i], size, postingsOutput);
// write vectors
bulkWriter.writeVectors(onHeapQuantizedVectors);
}
Expand Down Expand Up @@ -233,12 +255,15 @@ LongValues buildAndWritePostingsLists(
quantizedVectorsInput,
fieldInfo.getVectorDimension()
);
DocIdsWriter docIdsWriter = new DocIdsWriter();
DiskBBQBulkWriter bulkWriter = new DiskBBQBulkWriter.OneBitDiskBBQBulkWriter(ES91OSQVectorsScorer.BULK_SIZE, postingsOutput);
final ByteBuffer buffer = ByteBuffer.allocate(fieldInfo.getVectorDimension() * Float.BYTES).order(ByteOrder.LITTLE_ENDIAN);
// write the max posting list size
postingsOutput.writeVInt(maxPostingListSize);
// write the posting lists
int[] docIds = null;
int[] docDeltas = null;
int[] clusterOrds = null;
DocIdsWriter idsWriter = new DocIdsWriter();
for (int c = 0; c < centroidSupplier.size(); c++) {
float[] centroid = centroidSupplier.centroid(c);
int[] cluster = assignmentsByCluster[c];
Expand All @@ -252,11 +277,29 @@ LongValues buildAndWritePostingsLists(
// write docIds
int size = cluster.length;
postingsOutput.writeVInt(size);
offHeapQuantizedVectors.reset(size, ord -> isOverspill[ord], ord -> cluster[ord]);
if (docIds == null || docIds.length < cluster.length) {
docIds = new int[cluster.length];
clusterOrds = new int[cluster.length];
docDeltas = new int[cluster.length];
}
for (int j = 0; j < size; j++) {
docIds[j] = floatVectorValues.ordToDoc(cluster[j]);
clusterOrds[j] = j;
}
final int[] finalDocs = docIds;
final int[] finalOrds = clusterOrds;
// sort cluster.buffer by docIds values, this way cluster ordinals are sorted by docIds
new IntSorter(clusterOrds, i -> finalDocs[i]).sort(0, size);
// encode doc deltas
for (int j = 0; j < size; j++) {
docDeltas[j] = j == 0 ? finalDocs[finalOrds[j]] : finalDocs[finalOrds[j]] - finalDocs[finalOrds[j - 1]];
}
final int[] finalDocDeltas = docDeltas;
offHeapQuantizedVectors.reset(size, ord -> isOverspill[finalOrds[ord]], ord -> cluster[finalOrds[ord]]);
// TODO we might want to consider putting the docIds in a separate file
// to aid with only having to fetch vectors from slower storage when they are required
// keeping them in the same file indicates we pull the entire file into cache
docIdsWriter.writeDocIds(j -> floatVectorValues.ordToDoc(cluster[j]), size, postingsOutput);
idsWriter.writeDocIds(i -> finalDocDeltas[i], size, postingsOutput);
// write vectors
bulkWriter.writeVectors(offHeapQuantizedVectors);
}
Expand Down Expand Up @@ -717,4 +760,37 @@ public void readQuantizedVector(int ord, boolean isOverspill) throws IOException
bitSum = quantizedVectorsInput.readShort();
}
}

private static class IntSorter extends IntroSorter {
int pivot = -1;
private final int[] arr;
private final IntToIntFunction func;

private IntSorter(int[] arr, IntToIntFunction func) {
this.arr = arr;
this.func = func;
}

@Override
protected void setPivot(int i) {
pivot = func.apply(arr[i]);
}

@Override
protected int comparePivot(int j) {
return Integer.compare(pivot, func.apply(arr[j]));
}

@Override
protected int compare(int a, int b) {
return Integer.compare(func.apply(arr[a]), func.apply(arr[b]));
}

@Override
protected void swap(int i, int j) {
final int tmp = arr[i];
arr[i] = arr[j];
arr[j] = tmp;
}
}
}
81 changes: 5 additions & 76 deletions server/src/main/java/org/elasticsearch/index/codec/vectors/DocIdsWriter.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,8 +23,6 @@
import org.apache.lucene.store.DataOutput;
import org.apache.lucene.store.IndexInput;
import org.apache.lucene.util.ArrayUtil;
import org.apache.lucene.util.DocBaseBitSetIterator;
import org.apache.lucene.util.FixedBitSet;
import org.apache.lucene.util.IntsRef;
import org.apache.lucene.util.LongsRef;
import org.apache.lucene.util.hnsw.IntToIntFunction;
Expand All @@ -42,7 +40,6 @@ final class DocIdsWriter {
public static final int DEFAULT_MAX_POINTS_IN_LEAF_NODE = 512;

private static final byte CONTINUOUS_IDS = (byte) -2;
private static final byte BITSET_IDS = (byte) -1;
private static final byte DELTA_BPV_16 = (byte) 16;
private static final byte BPV_21 = (byte) 21;
private static final byte BPV_24 = (byte) 24;
Expand Down Expand Up @@ -92,21 +89,11 @@ void writeDocIds(IntToIntFunction docIds, int count, DataOutput out) throws IOEx
}

int min2max = max - min + 1;
if (strictlySorted) {
if (min2max == count) {
// continuous ids, typically happens when segment is sorted
out.writeByte(CONTINUOUS_IDS);
out.writeVInt(docIds.apply(0));
return;
} else if (min2max <= (count << 4)) {
assert min2max > count : "min2max: " + min2max + ", count: " + count;
// Only trigger bitset optimization when max - min + 1 <= 16 * count in order to avoid
// expanding too much storage.
// A field with lower cardinality will have higher probability to trigger this optimization.
out.writeByte(BITSET_IDS);
writeIdsAsBitSet(docIds, count, out);
return;
}
if (strictlySorted && min2max == count) {
// continuous ids, typically happens when segment is sorted
out.writeByte(CONTINUOUS_IDS);
out.writeVInt(docIds.apply(0));
return;
}

if (min2max <= 0xFFFF) {
Expand Down Expand Up @@ -180,38 +167,6 @@ void writeDocIds(IntToIntFunction docIds, int count, DataOutput out) throws IOEx
}
}

private static void writeIdsAsBitSet(IntToIntFunction docIds, int count, DataOutput out) throws IOException {
int min = docIds.apply(0);
int max = docIds.apply(count - 1);

final int offsetWords = min >> 6;
final int offsetBits = offsetWords << 6;
final int totalWordCount = FixedBitSet.bits2words(max - offsetBits + 1);
long currentWord = 0;
int currentWordIndex = 0;

out.writeVInt(offsetWords);
out.writeVInt(totalWordCount);
// build bit set streaming
for (int i = 0; i < count; i++) {
final int index = docIds.apply(i) - offsetBits;
final int nextWordIndex = index >> 6;
assert currentWordIndex <= nextWordIndex;
if (currentWordIndex < nextWordIndex) {
out.writeLong(currentWord);
currentWord = 0L;
currentWordIndex++;
while (currentWordIndex < nextWordIndex) {
currentWordIndex++;
out.writeLong(0L);
}
}
currentWord |= 1L << index;
}
out.writeLong(currentWord);
assert currentWordIndex + 1 == totalWordCount;
}

/** Read {@code count} integers into {@code docIDs}. */
void readInts(IndexInput in, int count, int[] docIDs) throws IOException {
if (count == 0) {
Expand All @@ -225,9 +180,6 @@ void readInts(IndexInput in, int count, int[] docIDs) throws IOException {
case CONTINUOUS_IDS:
readContinuousIds(in, count, docIDs);
break;
case BITSET_IDS:
readBitSet(in, count, docIDs);
break;
case DELTA_BPV_16:
readDelta16(in, count, docIDs);
break;
Expand All @@ -245,36 +197,13 @@ void readInts(IndexInput in, int count, int[] docIDs) throws IOException {
}
}

private DocIdSetIterator readBitSetIterator(IndexInput in, int count) throws IOException {
int offsetWords = in.readVInt();
int longLen = in.readVInt();
scratchLongs.longs = ArrayUtil.growNoCopy(scratchLongs.longs, longLen);
in.readLongs(scratchLongs.longs, 0, longLen);
// make ghost bits clear for FixedBitSet.
if (longLen < scratchLongs.length) {
Arrays.fill(scratchLongs.longs, longLen, scratchLongs.longs.length, 0);
}
scratchLongs.length = longLen;
FixedBitSet bitSet = new FixedBitSet(scratchLongs.longs, longLen << 6);
return new DocBaseBitSetIterator(bitSet, count, offsetWords << 6);
}

private static void readContinuousIds(IndexInput in, int count, int[] docIDs) throws IOException {
int start = in.readVInt();
for (int i = 0; i < count; i++) {
docIDs[i] = start + i;
}
}

private void readBitSet(IndexInput in, int count, int[] docIDs) throws IOException {
DocIdSetIterator iterator = readBitSetIterator(in, count);
int docId, pos = 0;
while ((docId = iterator.nextDoc()) != DocIdSetIterator.NO_MORE_DOCS) {
docIDs[pos++] = docId;
}
assert pos == count : "pos: " + pos + ", count: " + count;
}

private static void readDelta16(IndexInput in, int count, int[] docIds) throws IOException {
final int min = in.readVInt();
final int half = count >> 1;
Expand Down