hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Erik Krogen (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HADOOP-15840) Slow RPC logger too spammy
Date Thu, 01 Nov 2018 18:19:00 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-15840?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16671959#comment-16671959
] 

Erik Krogen commented on HADOOP-15840:
--------------------------------------

Using microsecond precision alone isn't sufficient to address the fact that 3 std dev can
produce too many events for many installations; this would be hundreds of events per second
for us. 4 std dev should be a reasonable amount, but it seems to me that this should be configurable,
as it may not be suitable for an installation with a fairly low QPS. 

> Slow RPC logger too spammy
> --------------------------
>
>                 Key: HADOOP-15840
>                 URL: https://issues.apache.org/jira/browse/HADOOP-15840
>             Project: Hadoop Common
>          Issue Type: Improvement
>    Affects Versions: 2.8.0, 2.7.4, 3.0.0-alpha1
>            Reporter: Wei-Chiu Chuang
>            Priority: Minor
>
> HADOOP-12325 added a capability where "slow" RPCs are logged in NN log when ipc.server.log.slow.rpc
is enabled.
> The "slow" RPCs are supposed to be those whose processing time is outside 3 standard
deviation, and it supposed to account for 0.3% of total RPCs.
> However, I found in practice, NN marks more than 1% of total RPCs as slow, and I've seen
RPCs whose processing time is 1ms declared slow too.
> {noformat}
> 2018-10-08 01:48:33,203 WARN org.apache.hadoop.ipc.Server: Slow RPC : sendHeartbeat took
1 milliseconds to process from client 10.17.199.16:56645
> 2018-10-08 01:48:33,219 WARN org.apache.hadoop.ipc.Server: Slow RPC : sendHeartbeat took
1 milliseconds to process from client 10.17.190.44:36435
> 2018-10-08 01:48:33,308 WARN org.apache.hadoop.ipc.Server: Slow RPC : sendHeartbeat took
1 milliseconds to process from client 10.17.190.43:56530
> {noformat}
> This is too many. 1% means NN spits hundreds slow RPCs per second on average in NN log.
> How about:
>  # use 4 stddev?
>  # use microsecond precision. Majority of RPCs takes less than 1 millisecond anyway.
It makes the stddev calculation imprecise. An RPC could be calculated to spend 1 millisecond
and be marked as "slow", since it starts from one millisecond and ends in the next millisecond.
>  
> [~anu] any thoughts?



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: common-issues-help@hadoop.apache.org


Mime
View raw message