[ https://issues.apache.org/jira/browse/KUDU-2217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16255995#comment-16255995 ]
Todd Lipcon commented on KUDU-2217: ----------------------------------- That's interesting. This implies that we have a config change op 1.3 which has its "new_config.index" set to 2 instead of 3. It seems, though, that whenever a pending op is added in AddPendingOperationUnlocked, we are forcing the new_config().index() to match the op index. Maybe we can add a DCHECK into the code that checks the above invariant when committing a change config and try looping this test? > Clarify on report about decreasing index of committed Raft config > ----------------------------------------------------------------- > > Key: KUDU-2217 > URL: https://issues.apache.org/jira/browse/KUDU-2217 > Project: Kudu > Issue Type: Task > Components: consensus, test > Reporter: Alexey Serbin > Assignee: Alexey Serbin > > Prior bf1fcb0df5f2fbb6d1557b6b932bc67045fff531 changelist, the > RaftConsensusNonVoterITest.PromoteAndDemote scenario exhibited flakiness when > running the binary built in DEBUG configuration. In addition to being flaky, > the test output some interesting error messages, e.g. it was something like > {noformat} > I1030 22:33:14.041613 12284 raft_consensus.cc:2456] T > 2abeb033beb040499f686713f41ba96c P 9e14a0231da149b2926039f26070025d [term 1 > FOLLOWER]: Committing config change with OpId 1.3: config changed from index > 3 to 2, ... > {noformat} > Unfortunately, I could not find the log from that run, but I remember how > [~mpercy] and I were looking at the message, wondering why committed index > went down from 3 to 2. > It would be nice to reproduce that and understand what is the reason behind. > It might be a regression. -- This message was sent by Atlassian JIRA (v6.4.14#64029)