[
https://issues.apache.org/jira/browse/HBASE-20939?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16559189#comment-16559189
]
stack commented on HBASE-20939:
-------------------------------
Your scenario would explain the WAL corruption that I'm seeing over in
HBASE-18152?
On the below....
1. We call suspendIfNotReady on a event, and it returns true so we need to wait.
2. The event has been waked up, and the procedure will be added back to the
scheduler.
3. A worker picks up the procedure and finishes it.
4. We finally throw ProcedureSuspendException and the ProcedureExecutor suspend
us and store the state in procedure store.
....I've been looking at this sequence presuming that after #1, we updated
store BEFORE suspend. Over in HBASE-18152, there is 70millis between the
*Starting* and suspend and the subsequent *Dispatch*...on another thread. In
the WAL on a later read, the two events show as flipped => Corruption.
> There will be race when we call suspendIfNotReady and then throw
> ProcedureSuspendedException
> --------------------------------------------------------------------------------------------
>
> Key: HBASE-20939
> URL: https://issues.apache.org/jira/browse/HBASE-20939
> Project: HBase
> Issue Type: Sub-task
> Reporter: Duo Zhang
> Assignee: Duo Zhang
> Priority: Major
> Attachments: HBASE-20939.patch
>
>
> This is very typical usage in our procedure implementation, for example, in
> AssignProcedure, we will call AM.queueAssign and then suspend ourselves to
> wait until the AM finish processing our assign request.
> But there could be races. Think of this:
> 1. We call suspendIfNotReady on a event, and it returns true so we need to
> wait.
> 2. The event has been waked up, and the procedure will be added back to the
> scheduler.
> 3. A worker picks up the procedure and finishes it.
> 4. We finally throw ProcedureSuspendException and the ProcedureExecutor
> suspend us and store the state in procedure store.
> So we have a half done procedure in the procedure store for ever... This may
> cause assertion when loading procedures. And maybe the worker can not finish
> the procedure as when suspending we need to restore some state, for example,
> add something to RootProcedureState. But anyway, it will still lead to
> assertion or other unexpected errors.
> And this can not be done by simply adding a lock in the procedure, as most
> works are done in the ProcedureExecutor after we throw
> ProcedureSuspendException.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
