[ 
https://issues.apache.org/jira/browse/HBASE-2751?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12880316#action_12880316
 ] 

Jonathan Gray commented on HBASE-2751:
--------------------------------------

I'm not sure about #1 but like #2.  In general, when you open up a region or a 
new storefile, you may need to serve requests from it.  Maybe there are some 
special circumstances this makes sense (cluster startup?) but other than that 
it might be hard to figure when it would make sense.  We're doing work to try 
to keep the time a region is offline as short as possible and that would add to 
it in the case that requests are pending.

Perhaps tracking at a StoreFile level the last access time and in a background 
thread you could iterate all regions/stores/storefiles to close file references 
older than a certain access time?  This would be more suggestive and if you 
touch all the data you'd open everything at once.  If you are thinking more 
explicit like maxOpenFiles we'd need some kind of LRU algo I guess.

> Consider closing StoreFiles sometimes
> -------------------------------------
>
>                 Key: HBASE-2751
>                 URL: https://issues.apache.org/jira/browse/HBASE-2751
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Jean-Daniel Cryans
>            Priority: Minor
>             Fix For: 0.21.0
>
>
> Having a lot of regions per region server could be considered harmless if 
> most of them aren't used, but that's not really true at the moment. We keep 
> all files opened all the time (except for rolled HLogs). I'm thinking of 2 
> solutions
>  # Lazy open the store files, or at least close them down after we read the 
> file info. Or we could do this for every file except the most recent one.
>  # Close files when they're not in use. We need some heuristic to determine 
> when is the best moment to declare that a file can be closed. 
> Both solutions go hand in hand, and I think it would be a huge gain in order 
> to lower the ulimit and xceivers-related issues.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to