[ https://issues.apache.org/jira/browse/CASSANDRA-9715?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14612485#comment-14612485 ]
Hazel Bobrins commented on CASSANDRA-9715: ------------------------------------------ Updated - sorry seem to have posted blank before. > Secondary index out of sync > --------------------------- > > Key: CASSANDRA-9715 > URL: https://issues.apache.org/jira/browse/CASSANDRA-9715 > Project: Cassandra > Issue Type: Bug > Components: Core > Environment: RHEL 6.2 2.6.32-220.13.1.el6.x86_64 > Reporter: Hazel Bobrins > > On 2.0.15 ( we moved from 2.08 hoping this problem would go away) we am > seeing intermittent issues where a secondary index is getting out of sync. > Set up is a 6 node cluster with 3 data centers, two nodes in each and with a > RF of 2 in each data centre. > So far I have been unable to reproduce this synthetically but have seen > multiple instances across all nodes within the cluster. > Data set is very small ~40K keys and >100MB of data. We add maybe 1000 > records a day, delete ~500 and update ~200. Not a very write based system. > Reads we can push out to ~2000/sec. > Writes are done at CL ALL and reads at ONE > All examples so far have been triggered when a record has been deleted and > then other added with the same index cardinality; I think it has also always > been the last record in the set which was deleted before the addition. > On a flushed keyspace a sstable2json export of the primary index shows all > records correctly, however, an export of the secondary index is missing the > records. > nodetool rebuild_index does not resolve the problem > Nether does a compact or repair > A select on the primary key at CL ALL also has no impact > However, a select at CL ALL on the secondary index does resolve the problem. > There is currently a none critical record which is out of the index on one of > our nodes. If another key is added with the same index cardinality it is > added to the index correctly. If this is then removed it once again returns > empty. > We have checked all the obvious OS bits and confirmed our time sync (ntp > based). > At DEBUG level we see nothing obvious wrong when adding/removing keys to the > above broken entry. > Due to the very intermittent nature of this problem is been impossible so far > to gather any DEBUG logs of it failing; we have also been unsuccessful so far > in reproducing this in out QA. > I know this is not much to go on, if there is anything we can provide to help > expand what might be the issue please let me know and we'll provide it asap. -- This message was sent by Atlassian JIRA (v6.3.4#6332)