[jira] [Comment Edited] (KAFKA-12713) Report "REAL" follower/consumer fetch latency

["Ming Liu (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-12713?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17331590#comment-17331590 ] 

Ming Liu edited comment on KAFKA-12713 at 4/27/21, 12:17 AM:
-------------------------------------------------------------

The idea is:
 #  Add waitTimeMs in FetchResponse
 #  In processResponseCallback() of handleFetchRequest,  set the waitTimeMs as the time spent in purgatory.
 #  In FetcherStats, we will add a new meter to track the fetch latency, by deduct the waitTimeMs from the latency. 

Also, in FetchLatency, we should also report a time called TotalEffectiveTime = TotalTime-RemoteTime. 

Let me know for any suggestion/feedback.  I like to propose a KIP on that change. 


was (Author: mingaliu):
The idea is:

0. Add waitTimeMs in Request()

1. In delayedOperation DelayedFetch class, add some code to track the actual wait time. 

2. In processResponseCallback() of handleFetchRequest, we can add additional parameter of waitTimeMs invoked from DelayedFetch.  It will set request.waitTimeMs.

3. In updateRequestMetrics() function, if waitTimeMs is not zero, we will deduct that out of RemoteTime and TotalTime.

Let me know for any suggestion/feedback.  I like to propose a KIP on that change. 

> Report "REAL" follower/consumer fetch latency
> ---------------------------------------------
>
>                 Key: KAFKA-12713
>                 URL: https://issues.apache.org/jira/browse/KAFKA-12713
>             Project: Kafka
>          Issue Type: Bug
>            Reporter: Ming Liu
>            Priority: Major
>
> The fetch latency is an important metrics to monitor for the cluster performance. With ACK=ALL, the produce latency is affected primarily by broker fetch latency.
> However, currently the reported fetch latency didn't reflect the true fetch latency because it sometimes need to stay in purgatory and wait for replica.fetch.wait.max.ms when data is not available. This greatly affect the real P50, P99 etc. 
> I like to propose a KIP to be able track the real fetch latency for both broker follower and consumer. 
>  



--
This message was sent by Atlassian Jira
(v8.3.4#803005)