[ 
https://issues.apache.org/jira/browse/HADOOP-12325?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Anu Engineer updated HADOOP-12325:
----------------------------------
    Attachment: HADOOP-12325.002.patch

Thanks for detailed review [~xyao] . I have attached a revised patch. Please 
see below for my detailed comments.

bq. 1. Do you miss updating all the caller of ProtobufRpcEngine.call() to pass 
receiveTime using Time.monotonicNow() instead of the Time.now()?

fixed, I have reverted to using Time.now() for in this patch.

bq. 2. Do we need update WritableRpcEngine.java class with logSlowRpcCalls()?

I could not find any place where we were using WritableRpcEngine for real, 
hence I did not make that change.


bq.  3. NIT: Can you put the magic number 1024 as final variable like 
fixed

bq. 4. Can you change the following from
fixed

bq. 5. NIT: Rpc -> RPC to be consistent
fixed

bq. Can you make the SleepRequestProto accepting a duration parameter instead 
of the fixed SLEEP_DURATION (1000ms)?
done

bq. 7. Is it possible to test with 1K fast calls instead of 10K calls to save 
test resources without affecting the results?

I had benchmarked these calls and even with 10K it is in milliseconds. The 
reason  I was making 10 K calls is to make sure that the test exercises the 
computation and the statistical significance properly.



> RPC Metrics : Add the ability track and log slow RPCs
> -----------------------------------------------------
>
>                 Key: HADOOP-12325
>                 URL: https://issues.apache.org/jira/browse/HADOOP-12325
>             Project: Hadoop Common
>          Issue Type: Improvement
>          Components: ipc, metrics
>    Affects Versions: 2.7.1
>            Reporter: Anu Engineer
>            Assignee: Anu Engineer
>         Attachments: HADOOP-12325.001.patch, HADOOP-12325.002.patch
>
>
> This JIRA proposes to add a counter called RpcSlowCalls and also a 
> configuration setting that allows users to log really slow RPCs.  Slow RPCs 
> are RPCs that fall at 99th percentile. This is useful to troubleshoot why 
> certain services like name node freezes under heavy load.



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

Reply via email to