Skip to content

Apache Lucene open-source search software

fix large p99 match request latency when purgeCache is running (#12801)

Davis Cook72dbfb7cf259
diff --git a/lucene/monitor/src/java/org/apache/lucene/monitor/WritableQueryIndex.java b/lucene/monitor/src/java/org/apache/lucene/monitor/WritableQueryIndex.java
index d71e3729fa5..980e1083ce8 100644
--- a/lucene/monitor/src/java/org/apache/lucene/monitor/WritableQueryIndex.java
+++ b/lucene/monitor/src/java/org/apache/lucene/monitor/WritableQueryIndex.java
@@ -273,25 +273,36 @@ class WritableQueryIndex extends QueryIndex {
// The purge takes the write lock when creating the register log, and then when swapping out
// the old query cache. Within the second write lock guard, the contents of the register log
// are added to the new query cache, and the register log itself is removed.
+ // The write lock must be taken inside the commit lock in order to prevent attempting to acquire
+ // the write lock while a commit is in progress. If a request to acquire the write lock was to
+ // be made while a commit was in progress, it would be forced to wait for the commit to give up
+ // the read lock. This is not a problem in itself, but the searches that want to acquire the
+ // read lock have to wait on the request to acquire the write lock in order to prevent
+ // starvation, which results in searches waiting on purge waiting on commit which is problematic
+ // since commit is slow and search cannot be slow.
final ConcurrentMap<String, QueryCacheEntry> newCache = new ConcurrentHashMap<>();
- purgeLock.writeLock().lock();
- try {
- purgeCache = new ConcurrentHashMap<>();
- } finally {
- purgeLock.writeLock().unlock();
+ synchronized (commitLock) {
+ purgeLock.writeLock().lock();
+ try {
+ purgeCache = new ConcurrentHashMap<>();
+ } finally {
+ purgeLock.writeLock().unlock();
+ }
}
populator.populateCacheWithIndex(newCache);
- purgeLock.writeLock().lock();
- try {
- newCache.putAll(purgeCache);
- purgeCache = null;
- queries = newCache;
- } finally {
- purgeLock.writeLock().unlock();
+ synchronized (commitLock) {
+ purgeLock.writeLock().lock();
+ try {
+ newCache.putAll(purgeCache);
+ purgeCache = null;
+ queries = newCache;
+ } finally {
+ purgeLock.writeLock().unlock();
+ }
}
}