[
https://issues.apache.org/jira/browse/HBASE-16755?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15892760#comment-15892760
]
Ashu Pachauri commented on HBASE-16755:
---------------------------------------
[~Apache9]
{quote}
And there is another minor reason why we set force to true is that, we select
regions based on their total memstore size. If we do not flush all the contents
in memstore, then the best candidate may not be the one which has the maximum
memstore size.
Maybe we could introduce new methods in FlushPolicy to find out the lfushable
size and select regions based on that value. Of course this is only a nice to
have, can do it later if anyone has interest.
{quote}
Makes sense.
bq. On the patch, as now we rely on the FlushPolicy to always return something
to flush, maybe we should add some checks in the code? We may introduce new
flush policies in the future. we need to make sure that they also follow the
rule.
A flush policy that returns nothing to flush (given that the decision whether a
region is flushable is made before querying the flush policy) sounds
impractical to use. But, to protect against such scenarios, we can either pass
down the "emergencyFlush" flag to the Region#flush which resorts to flushing
all store (if flush policy returns none) in case of an emergency flush. Another
option could be to return FlushResult.Result.CANNOT_FLUSH from
HRegion#flushCache if flush policy returns no stores to flush for the region
and the caller (MemstoreFlusher) can decide to retry with forceFlushAllStores
under certain circumstances.
> Honor flush policy under global memstore pressure
> -------------------------------------------------
>
> Key: HBASE-16755
> URL: https://issues.apache.org/jira/browse/HBASE-16755
> Project: HBase
> Issue Type: Improvement
> Components: regionserver
> Reporter: Ashu Pachauri
> Assignee: Ashu Pachauri
> Fix For: 1.3.1
>
> Attachments: HBASE-16755.v0.patch
>
>
> When global memstore reaches the low water mark, we pick the best flushable
> region and flush all column families for it. This is a suboptimal approach in
> the sense that it leads to an unnecessarily high file creation rate and IO
> amplification due to compactions. We should still try to honor the underlying
> FlushPolicy.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
