[ https://issues.apache.org/jira/browse/HBASE-21351?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16657837#comment-16657837 ]
Duo Zhang commented on HBASE-21351: ----------------------------------- But first I need to revisit the rollback logic to know how do we make sure that there is only one PE worker which executes the rollback. The code of setRollback seems like a lock, but it also checks other things so I need to read it more carefully... > The force update thread may have race with PE worker when the procedure is > rolling back > --------------------------------------------------------------------------------------- > > Key: HBASE-21351 > URL: https://issues.apache.org/jira/browse/HBASE-21351 > Project: HBase > Issue Type: Sub-task > Components: proc-v2 > Reporter: Duo Zhang > Priority: Major > Fix For: 3.0.0, 2.2.0, 2.1.1, 2.0.3 > > > We will acquire the procExecutionLock for a procedure when force updating its > state to prevent race with PE worker, but this does not work then the > procedure is rolling back. > If a procedure is failed, we will mark the root procedure stack as FAILED, > and then start to rollback the whole procedure stack. We will pop every > procedure in the stack and try to rollback them. So we may change the state > of a procedure without holding its procExecutionLock when rolling back. > This means we may persist an intermediate state of a procedure and cause > corruption when loading procedures. -- This message was sent by Atlassian JIRA (v7.6.3#76005)