[jira] [Comment Edited] (HBASE-21811) region can be opened on two servers due to race condition with procedures and server reports

["Duo Zhang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-21811?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16758043#comment-16758043 ] 

Duo Zhang edited comment on HBASE-21811 at 2/1/19 7:32 AM:
-----------------------------------------------------------

Ping [~zghaobac], I think this should be blocker for 2.2.0. In general, there is no way to interrupt an ongoing executeProcedures call if it is stuck there, so the problem here could happen.

I think the problem here is that, the Open/CloseRegionProcedure has already been woken up by SCP, and then we have schedule a new OpenRegionProcedure, and suspend the ProedureEvent in the RegionStateNode, and then the remoteCallFailed method for the completed Open/CloseRegionProcedure is called again and we will wake up the new OpenRegionProcedure which is waiting on the ProcedureEvent, and this will lead to a double assign.

So here we need to know that we should not wake up the procedure. But I wonder if we have a better way to determine that the wake up is stale other than checking for ServerName... Maybe we should check whether we are actually waiting on the event before waking it up...


was (Author: apache9):
Ping [~zghaobac], I think this should be blocker for 2.2.0. In general, there is no way to interrupt an ongoing executeProcedures call if it is stuck there, so the problem here could happen.

I think the problem here is that, the Open/CloseRegionProcedure has already been woken up by SCP, and then we have schedule a new OpenRegionProcedure, and suspend the ProedureEvent in the RegionStateNode, and then the remoteCallFailed method for the completed Open/CloseRegionProcedure has been called again and we will wake up the new OpenRegionProcedure which is waiting on the ProcedureEvent, and this will lead to a double assign.

So here we need to know that we should not wake up the procedure. But I wonder if we have a better way to determine that the wake up is stale other than checking for ServerName... Maybe we should check whether we are actually waiting on the event before waking it up...

> region can be opened on two servers due to race condition with procedures and server reports
> --------------------------------------------------------------------------------------------
>
>                 Key: HBASE-21811
>                 URL: https://issues.apache.org/jira/browse/HBASE-21811
>             Project: HBase
>          Issue Type: Bug
>          Components: amv2
>    Affects Versions: 3.0.0
>            Reporter: Sergey Shelukhin
>            Assignee: Sergey Shelukhin
>            Priority: Blocker
>             Fix For: 3.0.0, 2.2.0
>
>         Attachments: HBASE-21811.patch
>
>
> Looks like the region server responses are being processed incorrectly in places allowing te region to be opened on two servers.
> * The region server report handling in procedures should check which server is reporting.
> * Also although I didn't check (and it isn't implicated in this bug), RS must check in OPEN that it's actually the correct RS master sent open to (w.r.t. start timestamp)
> This was previosuly "mitigated" by master killing the RS with incorrect reports, but due to race conditions with reports and assignment the kill was replaced with a warning, so now this condition persists.
> Regardless, the kill approach is not a good fix because there's still a window when a region can be opened on two servers.
> A region is being opened by server_48c. The server dies, and we process the retry correctly (retry=3 because 2 previous similar open failures were processed correctly).
> We start opening it on server_1aa now.
> {noformat}
> 2019-01-28 18:12:09,862 INFO  [KeepAlivePEWorker-104] assignment.RegionStateStore: pid=4915 updating hbase:meta row=8be2a423b16471b9417f0f7de04281c6, regionState=ABNORMALLY_CLOSED
> 2019-01-28 18:12:09,862 INFO  [KeepAlivePEWorker-104] procedure.ServerCrashProcedure: pid=11944, state=RUNNABLE:SERVER_CRASH_ASSIGN, hasLock=true; ServerCrashProcedure server=server_48c,17020,1548726406632, splitWal=true, meta=false found RIT pid=4915, ppid=7, state=WAITING:REGION_STATE_TRANSITION_CONFIRM_OPENED, hasLock=true; TransitRegionStateProcedure table=table, region=8be2a423b16471b9417f0f7de04281c6, ASSIGN; rit=OPENING, location=server_48c,17020,1548726406632, table=table, region=8be2a423b16471b9417f0f7de04281c6
> 2019-01-28 18:12:10,778 INFO  [KeepAlivePEWorker-80] assignment.TransitRegionStateProcedure: Retry=3 of max=2147483647; pid=4915, ppid=7, state=RUNNABLE:REGION_STATE_TRANSITION_CONFIRM_OPENED, hasLock=true; TransitRegionStateProcedure table=table, region=8be2a423b16471b9417f0f7de04281c6, ASSIGN; rit=ABNORMALLY_CLOSED, location=null
> ...
> 2019-01-28 18:12:10,902 INFO  [KeepAlivePEWorker-80] assignment.TransitRegionStateProcedure: Starting pid=4915, ppid=7, state=RUNNABLE:REGION_STATE_TRANSITION_GET_ASSIGN_CANDIDATE, hasLock=true; TransitRegionStateProcedure table=table, region=8be2a423b16471b9417f0f7de04281c6, ASSIGN; rit=ABNORMALLY_CLOSED, location=null; forceNewPlan=true, retain=false
> 2019-01-28 18:12:11,114 INFO  [PEWorker-7] assignment.RegionStateStore: pid=4915 updating hbase:meta row=8be2a423b16471b9417f0f7de04281c6, regionState=OPENING, regionLocation=server_1aa,17020,1548727658713
> {noformat}
> However, we get the remote procedure failure from 48c after we've already started that.
> It actually tried to open on the restarted RS, which makes me wonder if this is safe also w.r.t. other races - what if RS already initialized and didn't error out?
> Need to check if we verify the start code expected by master on RS when opening.
> {noformat}
> 2019-01-28 18:12:12,179 WARN  [RSProcedureDispatcher-pool4-t362] assignment.RegionRemoteProcedureBase: The remote operation pid=11050, ppid=4915, state=SUCCESS, hasLock=false; org.apache.hadoop.hbase.master.assignment.OpenRegionProcedure for region {ENCODED => 8be2a423b16471b9417f0f7de04281c6 ... to server server_48c,17020,1548726406632 failed
> org.apache.hadoop.hbase.ipc.ServerNotRunningYetException: org.apache.hadoop.hbase.ipc.ServerNotRunningYetException: Server server_48c,17020,1548727752747 is not running yet
> 2019-01-28 18:12:12,179 WARN  [RSProcedureDispatcher-pool4-t362] procedure.RSProcedureDispatcher: server server_48c,17020,1548726406632 is not up for a while; try a new one
> {noformat}
> Without any other reason (at least logged), the RIT immediately retries again and chooses a new candidate. It then retries again and goes to the new 48c, but that's unrelated.
> {noformat}
> 2019-01-28 18:12:12,289 INFO  [KeepAlivePEWorker-100] assignment.TransitRegionStateProcedure: Retry=4 of max=2147483647; pid=4915, ppid=7, state=RUNNABLE:REGION_STATE_TRANSITION_CONFIRM_OPENED, hasLock=true; TransitRegionStateProcedure table=table, region=8be2a423b16471b9417f0f7de04281c6, ASSIGN; rit=OPENING, location=server_1aa,17020,1548727658713
> ...
> 2019-01-28 18:12:12,668 INFO  [PEWorker-9] assignment.RegionStateStore: pid=4915 updating hbase:meta row=8be2a423b16471b9417f0f7de04281c6, regionState=OPENING, regionLocation=server_fd3,17020,1548727536972
> ....
> 2019-01-28 18:26:29,480 INFO  [KeepAlivePEWorker-154] assignment.RegionStateStore: pid=4915 updating hbase:meta row=8be2a423b16471b9417f0f7de04281c6, regionState=OPENING, regionLocation=server_48c,17020,1548727752747
> {noformat}
> What does happen though is that 1aa, that never got a chance to respond at the time that the RIT erroneously retried above, finishes opening the region - which master ignores
> {noformat}
> 2019-01-28 18:27:42,424 WARN  [RpcServer.default.FPBQ.Fifo.handler=29,queue=4,port=17000] assignment.TransitRegionStateProcedure: Received report OPENED transition from server_1aa,17020,1548727658713 for rit=OPENING, location=server_48c,17020,1548727752747, table=table, region=8be2a423b16471b9417f0f7de04281c6, pid=4915 but the region is not on it, should be a retry, ignore
> {noformat}
> And starts spamming these warnings until finally the region is open in two places.
> {noformat}
> 2019-01-28 18:27:45,500 WARN  [RpcServer.default.FPBQ.Fifo.handler=42,queue=2,port=17000] assignment.AssignmentManager: rit=OPENING, location=server_48c,17020,1548727752747, table=table, region=8be2a423b16471b9417f0f7de04281c6 reported OPEN on server=server_1aa,17020,1548727658713 but state has otherwise
> ...
> 2019-01-28 18:51:21,197 WARN  [RpcServer.default.FPBQ.Fifo.handler=10,queue=0,port=17000] assignment.AssignmentManager: rit=OPEN, location=server_48c,17020,1548727752747, table=table, region=8be2a423b16471b9417f0f7de04281c6 reported OPEN on server=server_1aa,17020,1548727658713 but state has otherwise
> {noformat}
> This can result in data loss.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)