hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "stack (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-11109) flush region sequence id may not be larger than all edits flushed
Date Sat, 03 May 2014 04:12:14 GMT

    [ https://issues.apache.org/jira/browse/HBASE-11109?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13988569#comment-13988569
] 

stack commented on HBASE-11109:
-------------------------------

Ok. Pass on my suggested optimization above.  I was going to make a hack appendNoSync just
to get back the highest sequence number outstanding on the WAL and after getting this, sync
to that point only (not all edits).   It might help in some light cases but would be no good
for real hotspotitng so put it aside as a hack.  Let me commit this patch because it 'correct'
if hadoopqa says it good and lets come back here after we mesh region sequenceid and mvcc.
More fun ahead.

> flush region sequence id may not be larger than all edits flushed
> -----------------------------------------------------------------
>
>                 Key: HBASE-11109
>                 URL: https://issues.apache.org/jira/browse/HBASE-11109
>             Project: HBase
>          Issue Type: Sub-task
>          Components: wal
>    Affects Versions: 0.99.0
>            Reporter: stack
>            Assignee: stack
>            Priority: Critical
>             Fix For: 0.99.0
>
>         Attachments: 11109.txt, 11109v2.txt
>
>
> This was found by [~jeffreyz]  See parent issue.  We have this issue since we put the
ring buffer/disrupter into the WAL (HBASE-10156).
> An edits region sequence id is set only after the edit has traversed the ring buffer.
 Flushing, we just up whatever the current region sequence id is.  Crossing the ring buffer
may take some time and is done by background threads.  The flusher may be taking the region
sequence id though edits have not yet made it across the ringbuffer: i.e. edits that are actually
scoped by the flush may have region sequence ids in excess of that of the flush sequence id
reported.
> The consequences are not exactly clear.  Would rather not have to find out so lets fix
this here.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message