Paul Brannan created KUDU-3794:
----------------------------------
Summary: KuduTableAlterer::Alter should clear meta cache for
AlreadyPresent
Key: KUDU-3794
URL: https://issues.apache.org/jira/browse/KUDU-3794
Project: Kudu
Issue Type: Bug
Components: client
Affects Versions: 1.18.1
Reporter: Paul Brannan
We believe we have hit a case where one or more clients ends up with stale
metadata after multiple clients attempt to add the same range partition
simultaneously. We have not been able to reproduce this in a small test case.
There may be a race condition in KuduTableAlterer::Alter:
* if partitioning is changed, the meta cache is cleared if AlterTable returns
OK (WaitForAlterTableToFinish can return any status, so long as the initial rpc
succeeds)
* If there are two clients trying to add the same range partition
simultaneously, one client will get OK and the other will get AlreadyPresent
* The client that gets AlreadyPresent does not clear its meta-cache and ends
up with stale data in its meta cache
Proposal: reorder KuduTableAlterer::Alter such that ClearCache() is called
before the first RETURN_NOT_OK.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
