jpountz commented on code in PR #12183:
URL: https://github.com/apache/lucene/pull/12183#discussion_r1235739084
##########
lucene/core/src/java/org/apache/lucene/index/TermStates.java:
##########
@@ -86,19 +92,58 @@ public TermStates(
* @param needsStats if {@code true} then all leaf contexts will be visited
up-front to collect
* term statistics. Otherwise, the {@link TermState} objects will be
built only when requested
*/
- public static TermStates build(IndexReaderContext context, Term term,
boolean needsStats)
+ public static TermStates build(
+ IndexSearcher indexSearcher, IndexReaderContext context, Term term,
boolean needsStats)
throws IOException {
assert context != null && context.isTopLevel;
final TermStates perReaderTermState = new TermStates(needsStats ? null :
term, context);
if (needsStats) {
- for (final LeafReaderContext ctx : context.leaves()) {
- // if (DEBUG) System.out.println(" r=" + leaves[i].reader);
- TermsEnum termsEnum = loadTermsEnum(ctx, term);
- if (termsEnum != null) {
- final TermState termState = termsEnum.termState();
- // if (DEBUG) System.out.println(" found");
- perReaderTermState.register(
- termState, ctx.ord, termsEnum.docFreq(),
termsEnum.totalTermFreq());
+ Executor executor = indexSearcher.getExecutor();
+ boolean isShutdown = false;
+ if (executor instanceof ExecutorService) {
+ isShutdown = ((ExecutorService) executor).isShutdown();
+ }
+ if (executor != null && isShutdown == false) {
+ // build term states concurrently
+ List<FutureTask<Integer>> tasks =
+ context.leaves().stream()
+ .map(
+ ctx ->
+ new FutureTask<>(
+ () -> {
+ TermsEnum termsEnum = loadTermsEnum(ctx, term);
+ if (termsEnum != null) {
+ final TermState termState =
termsEnum.termState();
+ perReaderTermState.register(
+ termState,
+ ctx.ord,
+ termsEnum.docFreq(),
+ termsEnum.totalTermFreq());
+ }
+ return 0;
+ }))
+ .toList();
+ for (FutureTask<Integer> task : tasks) {
+ executor.execute(task);
Review Comment:
It would make more sense to me to parallelize at the segment level than at
the slice level indeed since vector search has a much higher per-segment cost
than per-document cost. I'm curious what others think about this, e.g.
@mikemccand ? To me slicing is really about things that have a per-document
cost, ie. running scorers / collectors.
--
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
