[
https://issues.apache.org/jira/browse/HBASE-22081?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16827407#comment-16827407
]
Sergey Shelukhin commented on HBASE-22081:
------------------------------------------
Before the patch, it is mere coincidence that while all the other stuff is
shutting down, the rpc that caused it has a chance to return.
Caller could get unlucky and stop RPC would fail because RPC server was
closed... now that we shut down RPC server first thing, it happens almost all
the time.
Added a small sleep before starting shutdown if it was triggered by and RPC
request 0_o Unfortunately it doesn't seem to be possible externally to wait for
RPC(s) responses to finish.
> master shutdown: close RpcServer and procWAL first thing
> --------------------------------------------------------
>
> Key: HBASE-22081
> URL: https://issues.apache.org/jira/browse/HBASE-22081
> Project: HBase
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Priority: Major
> Attachments: HBASE-22081.01.patch, HBASE-22081.patch
>
>
> I had a master get stuck due to HBASE-22079 and noticed it was logging RS
> abort messages during shutdown.
> [~bahramch] found some issues where messages are processed by old master
> during shutdown due to a race condition in RS cache (or it could also happen
> due to a network race).
> Previously I found some bug where SCP was created during master shutdown that
> had incorrect state (because some structures already got cleaned).
> I think before master fencing is implemented we can at least make these
> issues much less likely by thinking about shutdown order.
> 1) First kill RCP server so we don't receive any more messages. There's no
> need to receive messages when we are shutting down. Server heartbeats could
> be impacted I guess, but I don't think they will be cause we currently only
> kill RS on ZK timeout.
> 2) Then do whatever cleanup we think is needed that requires proc wal.
> 3) Then close proc WAL so no errant threads can create more procs.
> 4) Then do whatever other cleanup.
> 5) Finally delete znode.
> Right now znode is deleted somewhat early I think, and RpcServer is closed
> very late.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
