[ 
https://issues.apache.org/jira/browse/FLINK-19160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17192726#comment-17192726
 ] 

Piotr Nowojski edited comment on FLINK-19160 at 9/9/20, 2:19 PM:
-----------------------------------------------------------------

-The problem is that only one usage of blocking call {{TaskMailbox#take}} out 
of two is marking blocked time as idle. Probably a simple solution is to move 
{{MailboxProcessor#idleTime}} metric accounting to {{TaskMailboxImpl}}.-

The key thing here is to differentiate if {{AsyncWaitOperato}} is causing the 
backpressure, or if it's being backpressured a the downstream task. In the 
first case, when it's waiting for {{.yield()}} in {{processElement()}} it 
should actually be accounted as busy time.

When it's whole task is backpressured, the upstream task won't be allowed to 
process more input, and the {{idleTime}} will be increased.

Also if mailbox actions (like {{AsyncWaitOperator}} producing output) are 
executed while output is not available (task is backpressured), they are very 
likely to block on requesting buffers from the network stack. But this is also 
accounted to {{idleTime}} by {{RecordWriter}} class.


was (Author: pnowojski):
The problem is that only one usage of blocking call {{TaskMailbox#take}} out of 
two is marking blocked time as idle. Probably a simple solution is to move 
{{MailboxProcessor#idleTime}} metric accounting to {{TaskMailboxImpl}}.

> When backpressured AsyncWaitOperator/ContinousFileReaderOperator are not idle
> -----------------------------------------------------------------------------
>
>                 Key: FLINK-19160
>                 URL: https://issues.apache.org/jira/browse/FLINK-19160
>             Project: Flink
>          Issue Type: Bug
>          Components: Runtime / Task
>    Affects Versions: 1.12.0, 1.11.2
>            Reporter: Konstantin Knauf
>            Priority: Major
>              Labels: beginner
>
> When AsyncWaitOperator/ContinousFileReaderOperator are backpressured, this 
> time is not account for as idle by the "idleTime" metric.



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

Reply via email to