Can you come up with a test that shows the problem ?
Consider opening a JIRA with anonymized master log, your test and proposed
solution (if you have one).
Cheers
On Thu, Sep 10, 2015 at 6:09 AM, zhou_shuaif...@sina.com <
zhou_shuaif...@sina.com> wrote:
> Hi all,
> I found a situation may cause region closed forever, and this
> situation happend usually on my cluster, version is 0.98.10, but 1.1.2 also
> have the problem:
> 1, master send region open to regionserver
> 2, rs open a handler do openregion
> 3, rs return resopnse to master
> 3, master not received the response, or timeout, send open region again
> 4, rs already opened the region
> 5, master processAlreadyOpenedRegion, update regionstate open in
> master memory
> 6, master received zk message region opened(for some reason late, eg:
> net work), and triger update regionstate open, but find that region already
> opened, ERROR!
> 7, master send close region, and region be closed forever.
>
> may be a solution is change processAlreadyOpenedRegion in class
> AssignmentManager:
>
>
> private void processAlreadyOpenedRegion(HRegionInfo region, ServerName sn) {
> // Remove region from in-memory transition and unassigned node from ZK
> // While trying to enable the table the regions of the table were
> // already enabled.
> LOG.debug("ALREADY_OPENED " + region.getRegionNameAsString()
> + " to " + sn);
> String encodedName = region.getEncodedName();
>
> /**
> * check region state in zk, if already opened, return; leave the
> regionStates work to zkStatus change to trigger.
> **/
>
>
>
> deleteNodeInStates(encodedName, "offline", sn,
> EventType.M_ZK_REGION_OFFLINE);
> regionStates.regionOnline(region, sn);
> }
> ------------------------------
> zhou_shuaif...@sina.com
>
