Skip to content

Apache Lucene open-source search software

Change `set.removeAll(list)` to `list.forEach(set::remove)` (#13052)

Dmitry Cherniachenko39c10a292912
diff --git a/lucene/CHANGES.txt b/lucene/CHANGES.txt
index f0a5dea07a8..2db6735e644 100644
--- a/lucene/CHANGES.txt
+++ b/lucene/CHANGES.txt
@@ -228,6 +228,8 @@ Optimizations
this will help boolean queries that consist of a mix OF FILTER clauses and
SHOULD clauses. (Adrien Grand)
+* GITHUB#13052: Avoid set.removeAll(list) O(n^2) performance trap in the UpgradeIndexMergePolicy (Dmitry Cherniachenko)
+
Bug Fixes
---------------------
* GITHUB#12866: Prevent extra similarity computation for single-level HNSW graphs. (Kaival Parikh)
diff --git a/lucene/core/src/java/org/apache/lucene/index/UpgradeIndexMergePolicy.java b/lucene/core/src/java/org/apache/lucene/index/UpgradeIndexMergePolicy.java
index 655d5747ec4..944fd780d8c 100644
--- a/lucene/core/src/java/org/apache/lucene/index/UpgradeIndexMergePolicy.java
+++ b/lucene/core/src/java/org/apache/lucene/index/UpgradeIndexMergePolicy.java
@@ -106,7 +106,11 @@ public class UpgradeIndexMergePolicy extends FilterMergePolicy {
// the resulting set contains all segments that are left over
// and will be merged to one additional segment:
for (final OneMerge om : spec.merges) {
- oldSegments.keySet().removeAll(om.segments);
+ // om.segments.forEach(::remove) is used here instead of oldSegments.keySet().removeAll()
+ // for performance reasons; when om.segments.size() == oldSegments.size()
+ // the AbstractSet#removeAll() implementation will iterate the set elements
+ // calling list.contains() for each of them, resulting in O(n^2) performance
+ om.segments.forEach(oldSegments::remove);
}
}