[ 
https://issues.apache.org/jira/browse/HBASE-12335?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14189410#comment-14189410
 ] 

Nick Dimiduk commented on HBASE-12335:
--------------------------------------

To summarize my findings, it looks like region replicas has a very nice impact 
on 99.99pct latency results. Overall the stdev for these runs is minimized, 
which means the feature results in a much more uniform client experience. 
What's also interesting is the 99pct impact. With replica enabled, this jumps 
to 11ms, which is basically the configured replica timeout trigger (10) + 1. 
All these numbers come from the bottom spreadsheet I linked earlier.

These statistics are summarizing the 14 individual clients invoked in one run 
of this test by averaging the results. Also note that this was 4 clients over 
1mm rows, so the 99.99 pct is only 25 data points. Probably these numbers are 
not indicative of a long-running service. All times are in ms.

||metic||without replicas||with replicas||
|stdev|55.8301|2.1845|
|mean|1.1777|1.3125|
|99pct|2.0538|11.1075|
|99.9pct|16.2626|16.2988|
|99.99pct|125.1336|28.5603|

> IntegrationTestRegionReplicaPerf is flaky
> -----------------------------------------
>
>                 Key: HBASE-12335
>                 URL: https://issues.apache.org/jira/browse/HBASE-12335
>             Project: HBase
>          Issue Type: Test
>          Components: test
>    Affects Versions: 0.99.0, 2.0.0
>            Reporter: Nick Dimiduk
>            Assignee: Nick Dimiduk
>             Fix For: 2.0.0, 0.99.2
>
>         Attachments: HBASE-12335.00-0.99.patch, HBASE-12335.00.patch, 
> HBASE-12335.00.patch, HBASE-12335.00.patch, HBASE-12335.01-0.99.patch, 
> HBASE-12335.01.patch, HBASE-12335.02.patch
>
>
> I find that this test often fails; the assertion that running with read 
> replicas should complete faster than without is usually false. I need to 
> investigate further as to why this is the case and how we should tune it.
> In the mean time, I'd like to change the test to assert instead on the 
> average of the stdev across all the test runs in each category. Meaning, 
> enabling this feature should reduce the overall latency variance experienced 
> by the client.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to