Skip to content

[core] Support batch primary-key vector search #8666

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
JingsongLi merged 2 commits into from
Jul 15, 2026
Merged

[core] Support batch primary-key vector search #8666

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e1e3e72
[core] Support batch primary-key vector search
JingsongLi Jul 15, 2026
a98f2c8
[spark] Support primary-key batch vector results in lateral search
JingsongLi Jul 15, 2026
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
114 changes: 113 additions & 1 deletion paimon-core/src/main/java/org/apache/paimon/index/pkvector/PkVectorAnnSegmentSearcher.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,7 @@
import org.apache.paimon.index.pk.PrimaryKeyIndexSourceFile;
import org.apache.paimon.index.pk.PrimaryKeyIndexSourceMeta;
import org.apache.paimon.options.Options;
import org.apache.paimon.predicate.BatchVectorSearch;
import org.apache.paimon.predicate.VectorSearch;
import org.apache.paimon.types.DataField;
import org.apache.paimon.utils.IOUtils;
Expand Down Expand Up @@ -177,7 +178,7 @@ CompletableFuture<List<PkVectorSearchResult>> searchAsync(
"Vector segment %s has no source metadata.",
segment.fileName());
if (segment.rowCount() == 0) {
return Collections.emptyList();
return CompletableFuture.completedFuture(Collections.emptyList());
}
GlobalIndexer indexer =
GlobalIndexer.create(segment.indexType(), vectorField, indexOptions);
Expand Down Expand Up @@ -231,6 +232,117 @@ CompletableFuture<List<PkVectorSearchResult>> searchAsync(
}
}

public List<List<PkVectorSearchResult>> searchBatch(
IndexFileMeta segment,
PrimaryKeyIndexSourceMeta sourceMeta,
float[][] queries,
int limit,
Map<String, DeletionVector> deletionVectors,
Set<String> activeSourceFiles,
Map<String, List<Range>> rowRangesByFile,
Map<String, String> searchOptions) {
return searchBatchAsync(
segment,
sourceMeta,
queries,
limit,
deletionVectors,
activeSourceFiles,
rowRangesByFile,
searchOptions)
.join();
}

CompletableFuture<List<List<PkVectorSearchResult>>> searchBatchAsync(
IndexFileMeta segment,
PrimaryKeyIndexSourceMeta sourceMeta,
float[][] queries,
int limit,
Map<String, DeletionVector> deletionVectors,
Set<String> activeSourceFiles,
Map<String, List<Range>> rowRangesByFile,
Map<String, String> searchOptions) {
checkArgument(queries != null && queries.length > 0, "Query vectors cannot be empty.");
checkArgument(limit > 0, "Vector search limit must be positive: %s.", limit);
GlobalIndexMeta globalIndexMeta = segment.globalIndexMeta();
checkArgument(
globalIndexMeta != null && globalIndexMeta.sourceMeta() != null,
"Vector segment %s has no source metadata.",
segment.fileName());
if (segment.rowCount() == 0) {
List<List<PkVectorSearchResult>> results = new ArrayList<>(queries.length);
for (int i = 0; i < queries.length; i++) {
results.add(Collections.emptyList());
}
return CompletableFuture.completedFuture(Collections.unmodifiableList(results));
}
GlobalIndexer indexer =
GlobalIndexer.create(segment.indexType(), vectorField, indexOptions);
checkArgument(
indexer instanceof VectorGlobalIndexer,
"Index algorithm %s does not implement VectorGlobalIndexer.",
segment.indexType());
String readerMetric =
VectorSearchMetric.normalize(((VectorGlobalIndexer) indexer).metric());
checkArgument(
metric.equals(readerMetric),
"ANN segment metric %s does not match index reader metric %s.",
metric,
readerMetric);

GlobalIndexIOMeta ioMeta =
new GlobalIndexIOMeta(
annSegmentFile.path(segment),
segment.fileSize(),
globalIndexMeta.indexMeta());
GlobalIndexReader reader =
indexer.createReader(
meta -> fileIO.newInputStream(meta.filePath()),
Collections.singletonList(ioMeta),
executor);
try {
BatchVectorSearch search =
new BatchVectorSearch(queries, limit, vectorField.name(), searchOptions);
RoaringNavigableMap64 liveRows =
liveRowPositions(
sourceMeta.sourceFiles(),
activeSourceFiles,
deletionVectors,
rowRangesByFile);
if (liveRows != null) {
search.withIncludeRowIds(liveRows);
}
return reader.visitBatchVectorSearch(search)
.whenComplete((ignored, error) -> IOUtils.closeQuietly(reader))
.thenApply(
scoredResults -> {
checkArgument(
scoredResults.size() == queries.length,
"ANN segment %s returned %s batch results for %s queries.",
segment.fileName(),
scoredResults.size(),
queries.length);
List<List<PkVectorSearchResult>> results =
new ArrayList<>(queries.length);
for (Optional<ScoredGlobalIndexResult> scoredResult :
scoredResults) {
results.add(
mapResults(
segment,
sourceMeta,
deletionVectors,
activeSourceFiles,
rowRangesByFile,
scoredResult));
}
return Collections.unmodifiableList(results);
});
} catch (RuntimeException | Error t) {
IOUtils.closeQuietly(reader);
throw t;
}
}

private List<PkVectorSearchResult> mapResults(
IndexFileMeta segment,
PrimaryKeyIndexSourceMeta sourceMeta,
Expand Down
79 changes: 57 additions & 22 deletions paimon-core/src/main/java/org/apache/paimon/index/pkvector/PkVectorExactSearcher.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,44 +43,79 @@ public static List<PkVectorSearchResult> search(
int limit,
LongPredicate excludedPosition)
throws IOException {
checkArgument(query.length == reader.dimension(), "Query vector dimension does not match.");
return searchBatch(
dataFileName,
reader,
new float[][] {query},
metric,
limit,
excludedPosition)
.get(0);
}

public static List<List<PkVectorSearchResult>> searchBatch(
String dataFileName,
PkVectorReader reader,
float[][] queries,
String metric,
int limit,
LongPredicate excludedPosition)
throws IOException {
checkArgument(queries != null && queries.length > 0, "Query vectors cannot be empty.");
checkArgument(limit > 0, "Vector search limit must be positive.");
checkArgument(
VectorSearchMetric.isSupported(metric),
"Unsupported vector distance metric: %s.",
metric);
metric = VectorSearchMetric.normalize(metric);
for (int i = 0; i < query.length; i++) {
checkArgument(
Float.isFinite(query[i]),
"Query vector element at position %s must be finite.",
i);
}

Comparator<PkVectorSearchResult> bestFirst =
Comparator.comparingDouble(PkVectorSearchResult::distance)
.thenComparingLong(PkVectorSearchResult::rowPosition);
PriorityQueue<PkVectorSearchResult> nearest =
new PriorityQueue<>(limit, bestFirst.reversed());
List<PriorityQueue<PkVectorSearchResult>> nearest = new ArrayList<>(queries.length);
for (float[] query : queries) {
validateQuery(query, reader.dimension());
nearest.add(new PriorityQueue<>(limit, bestFirst.reversed()));
}

float[] vector = new float[reader.dimension()];
for (long position = 0; position < reader.rowCount(); position++) {
if (!reader.readNextVector(vector) || excludedPosition.test(position)) {
continue;
}
PkVectorSearchResult candidate =
new PkVectorSearchResult(
dataFileName,
position,
VectorSearchMetric.computeDistance(query, vector, metric));
if (nearest.size() < limit) {
nearest.add(candidate);
} else if (bestFirst.compare(candidate, nearest.peek()) < 0) {
nearest.poll();
nearest.add(candidate);
for (int i = 0; i < queries.length; i++) {
PkVectorSearchResult candidate =
new PkVectorSearchResult(
dataFileName,
position,
VectorSearchMetric.computeDistance(queries[i], vector, metric));
PriorityQueue<PkVectorSearchResult> queryNearest = nearest.get(i);
if (queryNearest.size() < limit) {
queryNearest.add(candidate);
} else if (bestFirst.compare(candidate, queryNearest.peek()) < 0) {
queryNearest.poll();
queryNearest.add(candidate);
}
}
}
List<PkVectorSearchResult> result = new ArrayList<>(nearest);
Collections.sort(result, bestFirst);
return Collections.unmodifiableList(result);

List<List<PkVectorSearchResult>> results = new ArrayList<>(queries.length);
for (PriorityQueue<PkVectorSearchResult> queryNearest : nearest) {
List<PkVectorSearchResult> result = new ArrayList<>(queryNearest);
Collections.sort(result, bestFirst);
results.add(Collections.unmodifiableList(result));
}
return Collections.unmodifiableList(results);
}

private static void validateQuery(float[] query, int dimension) {
checkArgument(query != null, "Query vector cannot be null.");
checkArgument(query.length == dimension, "Query vector dimension does not match.");
for (int i = 0; i < query.length; i++) {
checkArgument(
Float.isFinite(query[i]),
"Query vector element at position %s must be finite.",
i);
}
}
}
Loading
Loading