[ 
https://issues.apache.org/jira/browse/SPARK-49788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17890168#comment-17890168
 ] 

Alessandro Bellina commented on SPARK-49788:
--------------------------------------------

It would be great to have more discussion on shuffle cleanup mechanisms. 

Could this be triggered by size as well? Say you have a bunch of small shuffle 
data laying around the cluster, does it really matter that X amount of time has 
passed (e.g. why remove the small shuffle)?  Conversely, if your aggregate size 
of shuffle is large/growing then it makes a lot of sense to actively figure out 
how to nuke that data given some sort of priority/timestamp of last accessed.

Could we have a mode where after a query finishes executing ("spark is done 
with your action") where we actively trigger removal of those shuffle blocks?

> Add spark.cleaner.ttl functionality for long lived jobs
> -------------------------------------------------------
>
>                 Key: SPARK-49788
>                 URL: https://issues.apache.org/jira/browse/SPARK-49788
>             Project: Spark
>          Issue Type: Improvement
>          Components: Spark Core
>    Affects Versions: 4.0.0
>            Reporter: Holden Karau
>            Assignee: Holden Karau
>            Priority: Major
>
> We should add a TTL (and maybe a threshold?) to clean shuffle files which 
> have been stored for greater than some fixed period of time. This would be 
> useful for long lived jobs with Spark Connect and Notebooks where items may 
> not go out-of-scope and be resident for longer than needed.
>  
> See SPARK-7689 which removed the original TTL cleaner.
>  
> To reduce the chance of confusion/difficulty to understand which was part of 
> the original reason for removing the TTL based cleaner we should (instead of 
> "raw" TTL) keep track of when it was accessed last to reset the counter + log 
> on removal.
>  



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

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to