[ 
https://issues.apache.org/jira/browse/HBASE-27223?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17568484#comment-17568484
 ] 

chenglei edited comment on HBASE-27223 at 7/19/22 12:12 PM:
------------------------------------------------------------

[~zhangduo],  yes , from the design doc for HBASE-26233, I can see your point, 
but from the code, seems that I could not find "if WAL.sync fails, we will 
abort the region server", and moreover, may be the {{AsyncFSWAL}} is not failed 
completely, just the recovering from the failed datanode write takes some more 
time and {{AsyncFSWAL.sync}} is timeout and throws {{TimeoutIOException}}, in 
this situation, seems that the region server is not aborted.  Either way, I 
think even if this inconsistency is short, if we can avoid it, we should avoid 
it.


was (Author: comnetwork):
[~zhangduo],  yes , from the design doc for HBASE-26233, I can see your point, 
but from the code, seems that I could not find "if WAL.sync fails, we will 
abort the region server", and moreover, may be the {{AsyncFSWAL}} is not failed 
completely, just the recovering from the failed datanode takes some more time 
and {{AsyncFSWAL.sync}} is timeout and throws {{TimeoutIOException}}, in this 
situation, seems that the region server is not aborted.  Either way, I think 
even if this inconsistency is short, if we can avoid it, we should avoid it.

> Avoid data inconsistent between primary and secondary replicas for the new 
> region replication framework
> -------------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-27223
>                 URL: https://issues.apache.org/jira/browse/HBASE-27223
>             Project: HBase
>          Issue Type: Bug
>          Components: read replicas
>    Affects Versions: 3.0.0-alpha-3
>            Reporter: chenglei
>            Priority: Major
>
> For the new region replication framework introduced by HBASE-26233, when we 
> write the {{Mutation}} on primary region replica , if 
> {{HRegion.doMiniBatchMutate}}  throws a exception, for example ,{{WAL.sync}} 
> throws a {{TimeoutIOException}} because sync timeout, the writing on primary 
> region replica is failed and we could not query the {{Mutation}} on the 
> primary region replica because of {{Mutation}} is not written to the 
> {{MemStore}}, but the replication to the second replica may be successful 
> because we always execute 
> {{MultiVersionConcurrencyControl.WriteEntry.runCompletionAction}} in 
> {{MultiVersionConcurrencyControl.complete}} , and we could get the 
> {{Mutation}} on the secondary region replica, this inconsistency should be  
> avoid even if it is short. 
> But for {{WALUtil.writeMarker}}, we should always replicate to the second 
> replica even though {{WAL.sync}} is failed, because we just want to sync 
> region state to secondary replica.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to