[jira] [Comment Edited] (KAFKA-13370) Offset commit failure percentage metric is not computed correctly (regression)

["Randall Hauch (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-13370?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17444615#comment-17444615 ] 

Randall Hauch edited comment on KAFKA-13370 at 11/16/21, 3:47 PM:
------------------------------------------------------------------

Thanks, [~showuon], creating the PR that fixes this issue. But, I think the best course of action here is actually to revert https://github.com/apache/kafka/pull/9642, for a few reasons:

# It was actually useful to have that boolean success parameter within the protected TaskMetricsGroup.recordCommit(...) helper method, since it allowed the WorkerTask.recordCommitSuccess(...) and recordCommitFailure(...) methods to call that one method with desired behavior.
# Rolling back is also more encapsulated and significantly easier to backport all the way back to the 2.8 branch, especially since we've significantly refactored the source connector offset commit logic only in 3.0 and later branches.

So I think I'm going to close your PR and revert #9642.

However, I think your unit test improvement to verify the expected metric values are very useful. Would you mind creating a new PR with those unit test improvements? Since this is a blocker issue for 3.1, it'd be great to do that quickly. If that's not feasible (or I don't hear back in the next few days), then let's create a new issue for the additional unit test metric verification and associate the PR with that new issue.

Thanks!


was (Author: rhauch):
Thanks, [~showuon], creating the PR that fixes this issue. But, I think the best course of action here is actually to revert #9642, for a few reasons:

# It was actually useful to have that boolean success parameter within the protected TaskMetricsGroup.recordCommit(...) helper method, since it allowed the WorkerTask.recordCommitSuccess(...) and recordCommitFailure(...) methods to call that one method with desired behavior.
# Rolling back is also more encapsulated and significantly easier to backport all the way back to the 2.8 branch, especially since we've significantly refactored the source connector offset commit logic only in 3.0 and later branches.

So I think I'm going to close your PR and revert #9642.

However, I think your unit test improvement to verify the expected metric values are very useful. Would you mind creating a new PR with those unit test improvements? Since this is a blocker issue for 3.1, it'd be great to do that quickly. If that's not feasible (or I don't hear back in the next few days), then let's create a new issue for the additional unit test metric verification and associate the PR with that new issue.

Thanks!

> Offset commit failure percentage metric is not computed correctly (regression)
> ------------------------------------------------------------------------------
>
>                 Key: KAFKA-13370
>                 URL: https://issues.apache.org/jira/browse/KAFKA-13370
>             Project: Kafka
>          Issue Type: Bug
>          Components: KafkaConnect, metrics
>    Affects Versions: 2.8.0
>         Environment: Confluent Platform Helm Chart (v6.2.0)
>            Reporter: Vincent Giroux
>            Assignee: Luke Chen
>            Priority: Blocker
>             Fix For: 3.1.0, 3.0.1
>
>
> There seems to have been a regression in the way the offset-commit-* metrics are calculated for *source* Kafka Connect connectors since version 2.8.0.
> Before this version, any timeout or interruption while trying to commit offsets for source connectors (e.g. MM2 MirrorSourceConnector) would get correctly flagged as an offset commit failure (i.e the *offset-commit-failure-percentage* metric ** would be non-zero). Since version 2.8.0, these errors are considered as successes.
> After digging through the code, the commit where this bug was introduced appears to be this one : [https://github.com/apache/kafka/commit/047ad654da7903f3903760b0e6a6a58648ca7715]
> I believe removing the boolean *success* argument in the *recordCommit* method of the *WorkerTask* class (argument deemed redundant because of the presence of the Throwable *error* argument) and only considering the presence of a non-null error to determine if a commit is a success or failure might be a mistake. This is because in the *commitOffsets* method of the *WorkerSourceTask* class, there are multiple cases where an exception object is either not available or is not passed to the *recordCommitFailure* method, e.g. :
>  * *TImeout #1* : [https://github.com/apache/kafka/blob/2.8/connect/runtime/src/main/java/org/apache/kafka/connect/runtime/WorkerSourceTask.java#L519] 
>  * *Timeout #2* : [https://github.com/apache/kafka/blob/2.8/connect/runtime/src/main/java/org/apache/kafka/connect/runtime/WorkerSourceTask.java#L584] 
>  * *Interruption* : [https://github.com/apache/kafka/blob/2.8/connect/runtime/src/main/java/org/apache/kafka/connect/runtime/WorkerSourceTask.java#L529] 
>  * *Unserializable offset* : [https://github.com/apache/kafka/blob/2.8/connect/runtime/src/main/java/org/apache/kafka/connect/runtime/WorkerSourceTask.java#L562] 
>  



--
This message was sent by Atlassian Jira
(v8.20.1#820001)