hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jonathan Hsieh (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-11646) Handle the MOB in compaction
Date Wed, 06 Aug 2014 03:48:12 GMT

    [ https://issues.apache.org/jira/browse/HBASE-11646?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14087205#comment-14087205

Jonathan Hsieh commented on HBASE-11646:

The most important thing is that hbase serves the correct data regardless of whether we move
data into or out of the mob write path.  My gut says we'd want to put a new copy of the mob
them back into normal hfiles if the threshold is changed -- having it only go one way and
being irreversible seems inconsistent.  Can you come up with a case where we wouldn't want
or expect the copy back into hfiles behavior?

> Handle the MOB in compaction
> ----------------------------
>                 Key: HBASE-11646
>                 URL: https://issues.apache.org/jira/browse/HBASE-11646
>             Project: HBase
>          Issue Type: Sub-task
>          Components: Compaction
>            Reporter: Jingcheng Du
>            Assignee: Jingcheng Du
> In the updated MOB design however, admins can set CF level thresholds that would force
cell values > the threshold to use the MOB write path instead of the traditional path.
 There are two cases where mobs need to interact with this threshold
> 1) How do we handle the case when the threshold size is changed?
> 2) Today, you can bulkload hfiles that contain MOBs.  These cells will work as normal
inside hbase.  Unfortunately the cells with MOBs in them will never benefit form the MOB write
> The proposal here is to modify compaction in mob enabled cf's such that the threshold
value is honored with compactions.  This handles case #1 -- elements that should be moved
out of the normal hfiles get 'compacted' into refs and mob hfiles, and values that should
be pulled into the cf get derefed and written out wholy in the compaction.  For case #2, we
can maintain the same behavior and compaction would move data into the mob writepath/lifecycle.

This message was sent by Atlassian JIRA

View raw message