[ 
https://issues.apache.org/jira/browse/HBASE-3048?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12930279#action_12930279
 ] 

Kannan Muthukkaruppan commented on HBASE-3048:
----------------------------------------------

Prakash reported that he was seeing really good benefit of this optimization 
for his "incr" use case (which essentially rewrites the same keys over and over 
again). So doing delete processing in minor compactions reduce the working set 
nicely, and also improves block cache efficiencies and so on.



> unify code for major/minor compactions
> --------------------------------------
>
>                 Key: HBASE-3048
>                 URL: https://issues.apache.org/jira/browse/HBASE-3048
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Kannan Muthukkaruppan
>            Assignee: Amitanand Aiyer
>             Fix For: 0.90.0
>
>         Attachments: HBASE-3048-0.90-v2.patch, HBASE-3048-v3.patch, 
> unify.patch
>
>
> Today minor compactions do not process deletes, purge old versions, etc. Only 
> major compactions do.  The rationale was probably to save CPU (?). We should 
> evaluate if major compaction logic indeed runs significantly slower.
> Unifying minor compactions to do the same thing as major compactions has 
> other advantages:
> * If the same keys are deleted/updated repeatedly, the fact that 
> deletes/overwrites are not processed during minor compaction makes each 
> subsequent minor compaction more expensive as the total amount of data keeps 
> growing.
> * We'll have fewer bugs if the logic is as symmetric as possible. Any bugs in 
> TTL enforcement, version enforcement, etc. could cause behavior to be 
> different after a major compaction. Keeping the same logic means these bugs 
> will get caught earlier.
> -
> Note: There will still need to be one difference in the two schemes, and that 
> has to do with delete markers. Any compaction which doesn't compact all files 
> will still need to leave delete markers.

-- 
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