[jira] [Commented] (FLINK-31769) Add percentiles to aggregated metrics

["Gyula Fora (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-31769?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17713204#comment-17713204 ] 

Gyula Fora commented on FLINK-31769:
------------------------------------

I think this would be a great addition if we can add it with small overhead. Maybe make it configurable whether the user wants these or not

> Add percentiles to aggregated metrics
> -------------------------------------
>
>                 Key: FLINK-31769
>                 URL: https://issues.apache.org/jira/browse/FLINK-31769
>             Project: Flink
>          Issue Type: Improvement
>          Components: Autoscaler, Runtime / Metrics
>            Reporter: Zhanghao Chen
>            Priority: Major
>         Attachments: image-2023-04-11-15-11-51-471.png
>
>
> *Background*
> Currently only min/avg/max of metrics are exposed via REST API. Flink Autoscaler relies on these aggregated metrics to make predictions, and the type of aggregation plays an import role. [FLINK-30652] Use max busytime instead of average to compute true processing rate - ASF JIRA (apache.org) suggests that using max aggregator instead of avg of busy time can handle data skew more robustly. However, we found that for large-scale jobs, using max aggregation may be too sensitive. As a result, the true processing rate is underestimated with severe turbulence.
> The graph below is the true processing rate estimated with different aggregators of a real production data transmission job with a parallelism of 750.
> !image-2023-04-11-15-11-51-471.png!
> *Proposal*
> Add percentiles (p50, p90, p99) to aggregated metrics. Apache common maths can be used for computing that.
> A follow up would be making Flink autoscaler make use of the new aggregators.
>  
>  
>  
>  



--
This message was sent by Atlassian Jira
(v8.20.10#820010)