[
https://issues.apache.org/jira/browse/HBASE-10344?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Feng Honghua updated HBASE-10344:
---------------------------------
Description:
By fixing [HBASE-10329|https://issues.apache.org/jira/browse/HBASE-10329], we
know it's possible an asyncer's writes could have been synced by other asyncer
before it starts to do the sync. We distinguish this case and handle
accordingly when writer is null, but only for when writer is null.
When writer is not null(the much more typical case), we also can check and
ignore the sync if asyncer's writes have been synced by other asyncer before
doing sync. Since sync is a quite heavy operation(together with the following
notifying notifier thread and log roll check) we can have some performance gain
by ignoring it for such case.
was:
By fixing [HBASE-10329|https://issues.apache.org/jira/browse/HBASE-10329], we
know it's possible an asyncer's writes have been synced by other asyncer before
it starts to do the sync. We distinguish this case and handle accordingly when
the writer is null.
But when writer is not null(most of the time), we also can check and ignore the
sync if asyncer's writes have been synced by other asyncer before doing sync.
Since sync is a quite heavy operation(together with following notifying
notifier thread and log roll check) we can have some performance gain by
ignoring it for such case.
> Improve writer performance by ignoring sync to hdfs when an asyncer's writes
> have been synced by other asyncer
> --------------------------------------------------------------------------------------------------------------
>
> Key: HBASE-10344
> URL: https://issues.apache.org/jira/browse/HBASE-10344
> Project: HBase
> Issue Type: Improvement
> Reporter: Feng Honghua
> Assignee: Feng Honghua
> Attachments: HBASE-10344-trunk_v0.patch
>
>
> By fixing [HBASE-10329|https://issues.apache.org/jira/browse/HBASE-10329], we
> know it's possible an asyncer's writes could have been synced by other
> asyncer before it starts to do the sync. We distinguish this case and handle
> accordingly when writer is null, but only for when writer is null.
> When writer is not null(the much more typical case), we also can check and
> ignore the sync if asyncer's writes have been synced by other asyncer before
> doing sync. Since sync is a quite heavy operation(together with the following
> notifying notifier thread and log roll check) we can have some performance
> gain by ignoring it for such case.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
