[
https://issues.apache.org/jira/browse/HBASE-10156?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13870292#comment-13870292
]
stack commented on HBASE-10156:
-------------------------------
Here are a few other notes meantime.
+ I tested w/ raw dfs append/sync. There is a bunch of perf that we could pick
up here still comparing raw append/sync to what we are doing at least in single
threaded case (IIRC, the difference was nearly 10x -- we have more to do
appending, it is true, but it looks like there is a bunch more we could do in
here).
+ Anything we can do about how long syncs are taking in dfsclient will help a
bunch.
+ It is crazy how much work we do on each append. The calls to CP, WAL
listeners, and pb'ing records on EACH append needs review (WALs should be pb
all the time -- we should not be converting from WALEdit to pb WALEdit).
+ We can add back a sort of 'deferred flush' by not having Handlers wait on
their future to complete. We could do it on a per-Put basis rather than for
the whole regionserver. Should be easy enough to do after this patch goes in.
> Fix up the HBASE-8755 slowdown when low contention
> --------------------------------------------------
>
> Key: HBASE-10156
> URL: https://issues.apache.org/jira/browse/HBASE-10156
> Project: HBase
> Issue Type: Sub-task
> Components: wal
> Reporter: stack
> Assignee: stack
> Attachments: 10156.txt, 10156v2.txt, 10156v3.txt, 10156v4.txt,
> 10156v5.txt, 10156v6.txt, Disrupting.java
>
>
> HBASE-8755 slows our writes when only a few clients. Fix.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
