Enable optimistic search to memory optimized search. (#2933)

* Added MMapByteVectorValues for FP16 native scoring in LuceneOnFaiss. (#2904)

Signed-off-by: Dooyong Kim <[email protected]>

* Enable optimistic search for LuceneOnFaiss.

Signed-off-by: Dooyong Kim <[email protected]>

* Added debug logging, measure execution times of 2nd search.

Signed-off-by: Dooyong Kim <[email protected]>

---------

Signed-off-by: Dooyong Kim <[email protected]>
(cherry picked from commit a95a7d0)

Author: 0ctopus13prime

build.gradle

@@ -482,6 +482,7 @@ def commonIntegTest(RestIntegTestTask task, project, integTestDependOnJniLib, op
         task.systemProperty 'cluster.debug', isDebuggingCluster
         // Set number of nodes system property to be used in tests
         task.systemProperty 'cluster.number_of_nodes', "${_numNodes}"
+
         // There seems to be an issue when running multi node run or integ tasks with unicast_hosts
         // not being written, the waitForAllConditions ensures it's written
         task.getClusters().forEach { cluster ->
@@ -541,6 +542,7 @@ def commonIntegTestClusters(OpenSearchCluster cluster, _numNodes){
             debugPort += 1
         }
     }
+
     cluster.systemProperty("java.library.path", "$rootDir/jni/build/release")
     final testSnapshotFolder = file("${buildDir}/testSnapshotFolder")
     testSnapshotFolder.mkdirs()
@@ -550,6 +552,8 @@ def commonIntegTestClusters(OpenSearchCluster cluster, _numNodes){
 
 testClusters.integTest {
     commonIntegTestClusters(it, _numNodes)
+    // Forcing optimistic search for testing
+    systemProperty 'mem_opt_srch.force_reenter', 'true'
 }
 
 testClusters.integTestRemoteIndexBuild {
@@ -562,6 +566,8 @@ testClusters.integTestRemoteIndexBuild {
     keystore 's3.client.default.access_key', "${System.getProperty("access_key")}"
     keystore 's3.client.default.secret_key', "${System.getProperty("secret_key")}"
     keystore 's3.client.default.session_token', "${System.getProperty("session_token")}"
+    // Forcing optimistic search for testing
+    systemProperty 'mem_opt_srch.force_reenter', 'true'
 }
 
 task integTestRemote(type: RestIntegTestTask) {

src/main/java/org/opensearch/knn/index/query/KNNQueryFactory.java

@@ -117,7 +117,8 @@ public static Query create(CreateQueryRequest createQueryRequest) {
                         .build();
             }
 
-            if (createQueryRequest.getRescoreContext().isPresent()
+            if (memoryOptimizedSearchEnabled
+                || createQueryRequest.getRescoreContext().isPresent()
                 || (ENGINES_SUPPORTING_NESTED_FIELDS.contains(createQueryRequest.getKnnEngine()) && expandNested)) {
                 return new NativeEngineKnnVectorQuery(knnQuery, QueryUtils.getInstance(), expandNested);
             }

src/main/java/org/opensearch/knn/index/query/KNNWeight.java

@@ -24,7 +24,6 @@
 import org.apache.lucene.util.BitSetIterator;
 import org.apache.lucene.util.Bits;
 import org.apache.lucene.util.FixedBitSet;
-import org.opensearch.common.Nullable;
 import org.opensearch.common.StopWatch;
 import org.opensearch.common.lucene.Lucene;
 import org.opensearch.knn.common.FieldInfoExtractor;
@@ -52,6 +51,9 @@
 import static org.opensearch.knn.common.KNNConstants.SPACE_TYPE;
 import static org.opensearch.knn.common.KNNConstants.VECTOR_DATA_TYPE_FIELD;
 
+import static org.opensearch.knn.profile.StopWatchUtils.startStopWatch;
+import static org.opensearch.knn.profile.StopWatchUtils.stopStopWatchAndLog;
+
 /**
  * {@link KNNWeight} serves as a template for implementing approximate nearest neighbor (ANN)
  * and radius search over a native index type, such as Faiss.
@@ -298,12 +300,13 @@ public PerLeafResult searchLeaf(LeafReaderContext context, int k) throws IOExcep
         final SegmentReader reader = Lucene.segmentReader(context.reader());
         final String segmentName = reader.getSegmentName();
 
-        StopWatch stopWatch = startStopWatch();
+        final StopWatch stopWatch = startStopWatch(log);
         final BitSet filterBitSet = getFilteredDocsBitSet(context);
-        stopStopWatchAndLog(stopWatch, "FilterBitSet creation", segmentName);
+        stopStopWatchAndLog(log, stopWatch, "FilterBitSet creation", knnQuery.getShardId(), segmentName, knnQuery.getField());
+
+        // Save its cardinality, as the cardinality calculation is expensive.
+        final int filterCardinality = filterBitSet.cardinality();
 
-        final int maxDoc = context.reader().maxDoc();
-        int filterCardinality = filterBitSet.cardinality();
         // We don't need to go to JNI layer if no documents are found which satisfy the filters
         // We should give this condition a deeper look that where it should be placed. For now I feel this is a good
         // place,
@@ -320,19 +323,19 @@ public PerLeafResult searchLeaf(LeafReaderContext context, int k) throws IOExcep
          * This improves the recall.
          */
         if (isFilteredExactSearchPreferred(filterCardinality)) {
-            TopDocs result = doExactSearch(context, new BitSetIterator(filterBitSet, filterCardinality), filterCardinality, k);
-            return new PerLeafResult(filterWeight == null ? null : filterBitSet, result);
+            final TopDocs result = doExactSearch(context, new BitSetIterator(filterBitSet, filterCardinality), filterCardinality, k);
+            return new PerLeafResult(
+                filterWeight == null ? null : filterBitSet,
+                filterCardinality,
+                result,
+                PerLeafResult.SearchMode.EXACT_SEARCH
+            );
         }
 
-        /*
-         * If filters match all docs in this segment, then null should be passed as filterBitSet
-         * so that it will not do a bitset look up in bottom search layer.
-         */
-        final BitSet annFilter = (filterWeight != null && filterCardinality == maxDoc) ? null : filterBitSet;
+        final StopWatch annStopWatch = startStopWatch(log);
+        final TopDocs topDocs = approximateSearch(context, filterBitSet, filterCardinality, k);
+        stopStopWatchAndLog(log, stopWatch, "ANN search", knnQuery.getShardId(), segmentName, knnQuery.getField());
 
-        StopWatch annStopWatch = startStopWatch();
-        final TopDocs topDocs = approximateSearch(context, annFilter, filterCardinality, k);
-        stopStopWatchAndLog(annStopWatch, "ANN search", segmentName);
         if (knnQuery.isExplain()) {
             knnExplanation.addLeafResult(context.id(), topDocs.scoreDocs.length);
         }
@@ -341,18 +344,21 @@ public PerLeafResult searchLeaf(LeafReaderContext context, int k) throws IOExcep
         // results less than K, though we have more than k filtered docs
         if (isExactSearchRequire(context, filterCardinality, topDocs.scoreDocs.length)) {
             final BitSetIterator docs = filterWeight != null ? new BitSetIterator(filterBitSet, filterCardinality) : null;
-            TopDocs result = doExactSearch(context, docs, filterCardinality, k);
-            return new PerLeafResult(filterWeight == null ? null : filterBitSet, result);
+            final TopDocs result = doExactSearch(context, docs, filterCardinality, k);
+            return new PerLeafResult(
+                filterWeight == null ? null : filterBitSet,
+                filterCardinality,
+                result,
+                PerLeafResult.SearchMode.EXACT_SEARCH
+            );
         }
-        return new PerLeafResult(filterWeight == null ? null : filterBitSet, topDocs);
-    }
 
-    private void stopStopWatchAndLog(@Nullable final StopWatch stopWatch, final String prefixMessage, String segmentName) {
-        if (log.isDebugEnabled() && stopWatch != null) {
-            stopWatch.stop();
-            final String logMessage = prefixMessage + " shard: [{}], segment: [{}], field: [{}], time in nanos:[{}] ";
-            log.debug(logMessage, knnQuery.getShardId(), segmentName, knnQuery.getField(), stopWatch.totalTime().nanos());
-        }
+        return new PerLeafResult(
+            filterWeight == null ? null : filterBitSet,
+            filterCardinality,
+            topDocs,
+            PerLeafResult.SearchMode.APPROXIMATE_SEARCH
+        );
     }
 
     protected BitSet getFilteredDocsBitSet(final LeafReaderContext ctx) throws IOException {
@@ -413,9 +419,33 @@ private TopDocs doExactSearch(
         return exactSearch(context, exactSearcherContextBuilder.build());
     }
 
-    protected TopDocs approximateSearch(final LeafReaderContext context, final BitSet filterIdsBitSet, final int cardinality, final int k)
-        throws IOException {
+    /**
+     * Performs an approximate nearest neighbor (ANN) search on the provided index segment.
+     * <p>
+     * This method prepares all necessary query metadata before triggering the actual ANN search.
+     * It extracts the {@code model_id} from field-level attributes if required, retrieves any
+     * quantization or auxiliary metadata associated with the vector field, and applies quantization
+     * to the query vector when applicable. After these preprocessing steps, it invokes
+     * {@code doANNSearch(LeafReaderContext, BitSet, int, int)} to execute the approximate search
+     * and obtain the top results.
+     *
+     * @param context the {@link LeafReaderContext} representing the current index segment
+     * @param filterIdsBitSet an optional {@link BitSet} indicating document IDs to include in the search;
+     *                        may be {@code null} if no filtering is required
+     * @param filterCardinality the number of documents included in {@code filterIdsBitSet};
+     *                          used to optimize search filtering
+     * @param k the number of nearest neighbors to retrieve
+     * @return a {@link TopDocs} object containing the top {@code k} approximate search results
+     * @throws IOException if an error occurs while reading index data or accessing vector fields
+     */
+    public TopDocs approximateSearch(
+        final LeafReaderContext context,
+        final BitSet filterIdsBitSet,
+        final int filterCardinality,
+        final int k
+    ) throws IOException {
         final SegmentReader reader = Lucene.segmentReader(context.reader());
+
         FieldInfo fieldInfo = FieldInfoExtractor.getFieldInfo(reader, knnQuery.getField());
 
         if (fieldInfo == null) {
@@ -465,6 +495,11 @@ protected TopDocs approximateSearch(final LeafReaderContext context, final BitSe
         // TODO: Change type of vector once more quantization methods are supported
         byte[] quantizedVector = maybeQuantizeVector(segmentLevelQuantizationInfo);
         float[] transformedVector = maybeTransformVector(segmentLevelQuantizationInfo, spaceType);
+        /*
+         * If filters match all docs in this segment, then null should be passed as filterBitSet
+         * so that it will not do a bitset look up in bottom search layer.
+         */
+        final BitSet annFilter = filterCardinality == context.reader().maxDoc() ? null : filterIdsBitSet;
 
         KNNCounter.GRAPH_QUERY_REQUESTS.increment();
         final TopDocs results = doANNSearch(
@@ -477,8 +512,8 @@ protected TopDocs approximateSearch(final LeafReaderContext context, final BitSe
             quantizedVector,
             transformedVector,
             modelId,
-            filterIdsBitSet,
-            cardinality,
+            annFilter,
+            filterCardinality,
             k
         );
 
@@ -553,10 +588,10 @@ protected void addExplainIfRequired(final TopDocs results, final KNNEngine knnEn
      */
     public TopDocs exactSearch(final LeafReaderContext leafReaderContext, final ExactSearcher.ExactSearcherContext exactSearcherContext)
         throws IOException {
-        StopWatch stopWatch = startStopWatch();
+        final StopWatch stopWatch = startStopWatch(log);
         TopDocs exactSearchResults = exactSearcher.searchLeaf(leafReaderContext, exactSearcherContext);
         final SegmentReader reader = Lucene.segmentReader(leafReaderContext.reader());
-        stopStopWatchAndLog(stopWatch, "Exact search", reader.getSegmentName());
+        stopStopWatchAndLog(log, stopWatch, "Exact search", knnQuery.getShardId(), reader.getSegmentName(), knnQuery.getField());
         return exactSearchResults;
     }
 
@@ -673,13 +708,6 @@ private boolean isMissingNativeEngineFiles(LeafReaderContext context) {
         return engineFiles.isEmpty();
     }
 
-    private StopWatch startStopWatch() {
-        if (log.isDebugEnabled()) {
-            return new StopWatch().start();
-        }
-        return null;
-    }
-
     protected int[] getParentIdsArray(final LeafReaderContext context) throws IOException {
         if (knnQuery.getParentsFilter() == null) {
             return null;