Vincent Poon created PHOENIX-5046:
-------------------------------------
Summary: Race condition in disabling an index can cause an index
to get out of sync
Key: PHOENIX-5046
URL: https://issues.apache.org/jira/browse/PHOENIX-5046
Project: Phoenix
Issue Type: Bug
Affects Versions: 4.14.1, 5.0.0
Reporter: Vincent Poon
Assume a row R at T0.
If two index updates for R at T1 and T2 fail, the index might get marked
disabled as of T2 due to a race condition. The partial rebuilder will will
then rebuild as of T2. Since T1 was never replayed, the index row at T0 is not
deleted, leaving an extra orphan row in the index.
This is because In MetaDataEndpointImpl#updateIndexState , we update the index
state without any rowlocking, so even though we take the min of the new disable
timestamp and the current disable timestamp, two concurrent requests can be in
a race condition and succeed in disabling the index with different timestamps.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
