Major Description
A recent change increased the likelihood for this issue to show up but it can already happen before since we are using the attributes map in the StoredFieldsFormat for quite some time. I found this issue due to a test failure on our CI:
13:11:56 [junit4] Suite: org.apache.lucene.index.TestIndexSorting 13:11:56 [junit4] 2> apr 05, 2019 8:11:53 AM com.carrotsearch.randomizedtesting.RandomizedRunner$QueueUncaughtExceptionsHandler uncaughtException 13:11:56 [junit4] 2> WARNING: Uncaught exception in thread: Thread[Thread-507,5,TGRP-TestIndexSorting] 13:11:56 [junit4] 2> java.util.ConcurrentModificationException 13:11:56 [junit4] 2> at __randomizedtesting.SeedInfo.seed([7C25B308F180203B]:0) 13:11:56 [junit4] 2> at java.util.HashMap$HashIterator.nextNode(HashMap.java:1442) 13:11:56 [junit4] 2> at java.util.HashMap$EntryIterator.next(HashMap.java:1476) 13:11:56 [junit4] 2> at java.util.HashMap$EntryIterator.next(HashMap.java:1474) 13:11:56 [junit4] 2> at java.util.AbstractMap.toString(AbstractMap.java:554) 13:11:56 [junit4] 2> at org.apache.lucene.index.SegmentInfo.toString(SegmentInfo.java:222) 13:11:56 [junit4] 2> at org.apache.lucene.index.SegmentCommitInfo.toString(SegmentCommitInfo.java:345) 13:11:56 [junit4] 2> at org.apache.lucene.index.SegmentCommitInfo.toString(SegmentCommitInfo.java:364) 13:11:56 [junit4] 2> at java.lang.String.valueOf(String.java:2994) 13:11:56 [junit4] 2> at java.lang.StringBuilder.append(StringBuilder.java:131) 13:11:56 [junit4] 2> at java.util.AbstractMap.toString(AbstractMap.java:557) 13:11:56 [junit4] 2> at java.util.Collections$UnmodifiableMap.toString(Collections.java:1493) 13:11:56 [junit4] 2> at java.lang.String.valueOf(String.java:2994) 13:11:56 [junit4] 2> at java.lang.StringBuilder.append(StringBuilder.java:131) 13:11:56 [junit4] 2> at org.apache.lucene.index.TieredMergePolicy.findForcedMerges(TieredMergePolicy.java:628) 13:11:56 [junit4] 2> at org.apache.lucene.index.IndexWriter.updatePendingMerges(IndexWriter.java:2181) 13:11:56 [junit4] 2> at org.apache.lucene.index.IndexWriter.maybeMerge(IndexWriter.java:2154) 13:11:56 [junit4] 2> at org.apache.lucene.index.IndexWriter.forceMerge(IndexWriter.java:1988) 13:11:56 [junit4] 2> at org.apache.lucene.index.IndexWriter.forceMerge(IndexWriter.java:1939) 13:11:56 [junit4] 2> at org.apache.lucene.index.TestIndexSorting$UpdateRunnable.run(TestIndexSorting.java:1851) 13:11:56 [junit4] 2> at java.lang.Thread.run(Thread.java:748) 13:11:56 [junit4] 2> 13:11:56 [junit4] 2> NOTE: reproduce with: ant test -Dtestcase=TestIndexSorting -Dtests.method=testConcurrentUpdates -Dtests.seed=7C25B308F180203B -Dtests.slow=true -Dtest
The issue is that we update the attributes map (also we similarly do the same for diagnostics but it's not necessarily causing the issue since the diagnostics map is never modified) during the merge process but access it in the merge policy when looking at running merges and there we call toString on SegmentCommitInfo which happens without any synchronization. This is technically unsafe publication but IW is a mess along those lines and real fixes would require significant changes.
Attachments
Issue Links
- links to
