[ 
https://issues.apache.org/jira/browse/FLINK-26560?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17503971#comment-17503971
 ] 

Yun Tang commented on FLINK-26560:
----------------------------------

Upgrade the priority to "Major".

Once FLINK-23399 could be resolved, we can have benchmark to know what overlap 
fraction threshold is better in many cases, and we can then help improve the 
performance of RocksDB rescaling.

> Make the threshold of the overlap fraction of incremental restoring 
> configurable
> --------------------------------------------------------------------------------
>
>                 Key: FLINK-26560
>                 URL: https://issues.apache.org/jira/browse/FLINK-26560
>             Project: Flink
>          Issue Type: Improvement
>          Components: Runtime / State Backends
>    Affects Versions: 1.15.0
>            Reporter: Yanfei Lei
>            Assignee: Yanfei Lei
>            Priority: Minor
>
> Currently, the threshold of the overlap fraction of incremental restoring 
> `OVERLAP_FRACTION_THRESHOLD` is a hard-coded, fixed value.
>  
> {code:java}
> public class RocksDBIncrementalCheckpointUtils {
>     /**
>      * The threshold of the overlap fraction of the handle's key-group range 
> with target key-group
>      * range to be an initial handle.
>      */
>     private static final double OVERLAP_FRACTION_THRESHOLD = 0.75;
> ...
> } {code}
>  
> `OVERLAP_FRACTION_THRESHOLD` is used to control how to restore a state 
> handle, different thresholds can affect the performance of restoring. The 
> behavior of deletion in restoring has been changed after FLINK-21321, the old 
> threshold no longer fits the current situation.
> To make it easier to modify the threshold according to different situations, 
> changing `OVERLAP_FRACTION_THRESHOLD` to be configurable is suggested.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)

Reply via email to