[
https://issues.apache.org/jira/browse/FLINK-5090?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15679596#comment-15679596
]
Xiaogang Shi commented on FLINK-5090:
-------------------------------------
In flink, the performance bottlenecks are usually caused by
1. the mismatched parallelism of the producer and the consumer operators.
2. the imbalanced load across the different tasks of the same operator
The metrics of all channels help a lot to figure out the two problems.
But the solution to the second problem usually needs modification to the
application logic.
The gate-wise metrics are sufficient to identify the first problem.
I think it requires few additional overheads (due to two input operators).
> Expose optionally detailed metrics about network queue lengths
> --------------------------------------------------------------
>
> Key: FLINK-5090
> URL: https://issues.apache.org/jira/browse/FLINK-5090
> Project: Flink
> Issue Type: New Feature
> Components: Network
> Affects Versions: 1.1.3
> Reporter: Stephan Ewen
> Assignee: Stephan Ewen
> Fix For: 1.2.0, 1.1.4
>
>
> For debugging purposes, it is important to have access to more detailed
> metrics about the length of network input and output queues.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
