Munagala V. Ramanath commented on APEXCORE-556:

The `AbstractFileInputOperator` can potentially benefit from this attribute; 
the scenario is as follows:
During recovery, `beginWindow()` invokes `replay()` which retrieves files and 
offsets for this window
from the `WindowDataManager`, opens the files and reads records starting from 
the beginning,
discarding them till we reach the offset -- in a tight loop (there is actually 
a JIRA for a bug with a
related PR open https://issues.apache.org/jira/browse/APEXMALHAR-2270).
For a large file, this can take many minutes and result in the operator being 
considered hung and restarted
causing the same problem all over again.

> New operator attribute - RECOVERY_WINDOW_COUNT
> ----------------------------------------------
>                 Key: APEXCORE-556
>                 URL: https://issues.apache.org/jira/browse/APEXCORE-556
>             Project: Apache Apex Core
>          Issue Type: Improvement
>            Reporter: Sandesh
> During recovery operators need more time, so it is better to have different 
> per operator attribute representing that behaviour.
> Note:
> 1. Input and output operators need to have a higher value for this attribute, 
> as that is where recovery times are higher.

This message was sent by Atlassian JIRA

Reply via email to