[
https://issues.apache.org/jira/browse/HBASE-17324?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15759919#comment-15759919
]
Jingcheng Du commented on HBASE-17324:
--------------------------------------
Yes, I think a flag works for it. Now we can use --hbase.client.write.buffer=X
to set the buffer size for each test.
But BufferedMutator uses backgroundFlushCommits(false) for every mutation if
the buffer is 0, slightly different with Table.put where
backgroundFlushCommits(true) is used.
I think we have to use Table instead of BufferedMutator if the buffer is set as
0.
> PE Write workload results are wrong
> -----------------------------------
>
> Key: HBASE-17324
> URL: https://issues.apache.org/jira/browse/HBASE-17324
> Project: HBase
> Issue Type: Bug
> Reporter: Appy
>
> [~jmspaggi] found this issue of write latencies being unreasonable.
> The reason is, we are using BufferedMutator with size 2MB, so most writes
> return very quick because they get buffered. That's giving us avg. latency
> like 39us. Needs fixing.
> {noformat}
> 16/12/14 12:46:04 INFO hbase.PerformanceEvaluation: Latency (us) :
> mean=39,37, min=2,00, max=27193408,00, stdDev=3443,01, 50th=2,00, 75th=2,00,
> 95th=3,00, 99th=5,00, 99.9th=9157,00, 99.99th=39685,59, 99.999th=160751,28
> 16/12/14 12:46:04 INFO hbase.PerformanceEvaluation: Num measures (latency) :
> 100000000
> 16/12/14 12:46:04 INFO hbase.PerformanceEvaluation: Mean = 39,37
> Min = 2,00
> Max = 27193408,00
> StdDev = 3443,01
> 50th = 2,00
> 75th = 2,00
> 95th = 3,00
> 99th = 5,00
> 99.9th = 9157,00
> 99.99th = 39685,59
> 99.999th = 160751,28
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
