[
https://issues.apache.org/jira/browse/HBASE-26960?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
chenglei updated HBASE-26960:
-----------------------------
Description:
Besides HBASE-26768, there is another case replication in
{{RegionReplicationSink}} would be suspend:
For {{RegionReplicationSink}}, when there is a replication error ,
{{RegionReplicationSink}} invokes {{MemStoreFlusher#requestFlush}} to request a
flush, and after receiving the {{FlushAction#START_FLUSH}} or
{{FlushAction#CANNOT_FLUSH}} flush marker, it would resume the replication. But
when {{MemStoreFlusher}} flushing, it invokes following method
{{HRegion.flushcache}} with the {{writeFlushRequestWalMarker}} set to false:
{code:java}
public FlushResultImpl flushcache(List<byte[]> families,
boolean writeFlushRequestWalMarker, FlushLifeCycleTracker tracker) throws
IOException {
}
{code}
When {{writeFlushRequestWalMarker}} is set to false, {{HRegion.flushcache}}
does not write the {{FlushAction#CANNOT_FLUSH}} flush marker to {{WAL}} when
the memstore is empty, so when there is a replication error when the memstore
is empty(eg. replicating the {{FlushAction#START_FLUSH}} or
{{FlushAction#COMMIT_FLUSH}} ), the replication may suspend until next memstore
flush,even though later there are user writes and it could replicate normally.
I simulate this problem in the PR , and for {{writeFlushRequestWalMarker}}
paramter, it is introduced by HBASE-11580 and just only determines whether
writing the {{FlushAction#CANNOT_FLUSH}} flush marker to WAL when the memstore
is empty, so I think for simplicity, we could set it to true always for
{{MemStoreFlusher}}.
was:
There is another case replication in {{RegionReplicationSink}} would be
suspend:
For {{RegionReplicationSink}}, when there is a replication error ,
{{RegionReplicationSink}} invokes {{MemStoreFlusher#requestFlush}} to request a
flush, and after receiving the {{FlushAction#START_FLUSH}} or
{{FlushAction#CANNOT_FLUSH}} flush marker, it would resume the replication. But
when {{MemStoreFlusher}} flushing, it invokes following method
{{HRegion.flushcache}} with the {{writeFlushRequestWalMarker}} set to false:
{code:java}
public FlushResultImpl flushcache(List<byte[]> families,
boolean writeFlushRequestWalMarker, FlushLifeCycleTracker tracker) throws
IOException {
}
{code}
When {{writeFlushRequestWalMarker}} is set to false, {{HRegion.flushcache}}
does not write the {{FlushAction#CANNOT_FLUSH}} flush marker to {{WAL}} when
the memstore is empty, so when there is a replication error when the memstore
is empty(eg. replicating the {{FlushAction#START_FLUSH}} or
{{FlushAction#COMMIT_FLUSH}} ), the replication may suspend until next memstore
flush,even though later there are user writes and it could replicate normally.
I simulate this problem in the PR , and for {{writeFlushRequestWalMarker}}
paramter, it is introduced by HBASE-11580 and just only determines whether
writing the {{FlushAction#CANNOT_FLUSH}} flush marker to WAL when the memstore
is empty, so I think for simplicity, we could set it to true always for
{{MemStoreFlusher}}.
> Another case for unnecessary replication suspending in RegionReplicationSink
> ----------------------------------------------------------------------------
>
> Key: HBASE-26960
> URL: https://issues.apache.org/jira/browse/HBASE-26960
> Project: HBase
> Issue Type: Bug
> Components: read replicas
> Affects Versions: 3.0.0-alpha-2
> Reporter: chenglei
> Assignee: chenglei
> Priority: Major
>
> Besides HBASE-26768, there is another case replication in
> {{RegionReplicationSink}} would be suspend:
> For {{RegionReplicationSink}}, when there is a replication error ,
> {{RegionReplicationSink}} invokes {{MemStoreFlusher#requestFlush}} to request
> a flush, and after receiving the {{FlushAction#START_FLUSH}} or
> {{FlushAction#CANNOT_FLUSH}} flush marker, it would resume the replication.
> But when {{MemStoreFlusher}} flushing, it invokes following method
> {{HRegion.flushcache}} with the {{writeFlushRequestWalMarker}} set to false:
> {code:java}
> public FlushResultImpl flushcache(List<byte[]> families,
> boolean writeFlushRequestWalMarker, FlushLifeCycleTracker tracker)
> throws IOException {
> }
> {code}
> When {{writeFlushRequestWalMarker}} is set to false, {{HRegion.flushcache}}
> does not write the {{FlushAction#CANNOT_FLUSH}} flush marker to {{WAL}} when
> the memstore is empty, so when there is a replication error when the memstore
> is empty(eg. replicating the {{FlushAction#START_FLUSH}} or
> {{FlushAction#COMMIT_FLUSH}} ), the replication may suspend until next
> memstore flush,even though later there are user writes and it could replicate
> normally.
> I simulate this problem in the PR , and for {{writeFlushRequestWalMarker}}
> paramter, it is introduced by HBASE-11580 and just only determines whether
> writing the {{FlushAction#CANNOT_FLUSH}} flush marker to WAL when the
> memstore is empty, so I think for simplicity, we could set it to true always
> for {{MemStoreFlusher}}.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
