[ https://issues.apache.org/jira/browse/HIVE-17511?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16165443#comment-16165443 ]
Sergey Shelukhin commented on HIVE-17511: ----------------------------------------- Frankly the only explanation that I can see is if the same ProcCacheChunk is returned twice from the object pool due to a bug (or returned TO the pool twice by a thread and then legitimately returned twice from the pool). It seems like lists from two racing threads are merged at an item while both threads are just straightforwardly uncompressing ORC CBs linearly from a 100% cache miss. At the same time, given that after getting the item it's initialized, I'd expect one of the lists to have a completely invalid item, whereas here looks like only one link is invalid while all lists are contiguous, without the item that would be overwritten. So it's really weird. Looking at it now. Looks like the ordering checks patch that could have made the error clearer is missing from this build, I'm backporting it for now. Pool has pretty good multi-threaded tests so not sure yet how can this happen. > Error while populating orc cache in llap > ---------------------------------------- > > Key: HIVE-17511 > URL: https://issues.apache.org/jira/browse/HIVE-17511 > Project: Hive > Issue Type: Bug > Components: ORC > Reporter: Ashutosh Chauhan > Assignee: Sergey Shelukhin > > Observed that while querying an error is thrown while loading cache in llap > daemons -- This message was sent by Atlassian JIRA (v6.4.14#64029)