You are viewing a plain text version of this content. The canonical link for it is here.

Posted to issues@flink.apache.org by "fanrui (Jira)" <ji...@apache.org> on 2021/01/14 10:32:00 UTC

[jira] [Commented] (FLINK-16444) Count the read/write/seek/next latency of RocksDB as metrics

    [ https://issues.apache.org/jira/browse/FLINK-16444?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17264751#comment-17264751 ] 

fanrui commented on FLINK-16444:
--------------------------------

+1, some latency metrics are very useful.

Use flink-benckmark to make a benchmark. When RocksDB Statistics is enabled (StatsLevel.ALL), the update and value performance of ValueState is reduced by more than 10%.

Is it better to do some sampling on the Flink side? For example: every 100 calls to the `ValueState.update` method, latency is collected once.

> Count the read/write/seek/next latency of RocksDB as metrics
> ------------------------------------------------------------
>
>                 Key: FLINK-16444
>                 URL: https://issues.apache.org/jira/browse/FLINK-16444
>             Project: Flink
>          Issue Type: Improvement
>          Components: Runtime / State Backends
>            Reporter: Yun Tang
>            Assignee: Yun Tang
>            Priority: Major
>             Fix For: 1.13.0
>
>
> Currently, user cannot know the read/write/seek/next latency of RocksDB, we could add these helpful metrics to know the overall state performance. To not affect the action performance much, we could introduce counter to only record the latency at interval of some actions.



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