[
https://issues.apache.org/jira/browse/HBASE-724?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12611133#action_12611133
]
Sebastien Rainville commented on HBASE-724:
-------------------------------------------
Forgot to mention: testdeleteall.patch is to be applied on branch-0.1. Also, it
doesn't fix anything... it just makes the problem visible. I still don't know
where the problem is exactly but I looked at the HStore.get() implementation
and it seems really odd to me:
1) the code doesn't look if the LAST version of the cell is deleted but instead
just skip those that are deleted and still returns the other ones
2) getFull() works but not get()... it suggests that get() is faulty
3) (this one is not directly related to this issue) get() returns the cells in
the order that it finds it... how can it be working? i.e. how does the cells
are returned in chronological order? I must just not understand the algorithm
but let's say there are multiple versions of the same cell spanned in several
mapfiles... all the mapfiles are read sequentially so what garanties that the
versions will be found in chronological order? are the mapfiles somehow read in
chronological order? and within a mapfile, are the versions of a cell in
chronological order as well?
> Deleted rows reappear after a HStore.flushCache() and/or
> HStore.compactStores()
> -------------------------------------------------------------------------------
>
> Key: HBASE-724
> URL: https://issues.apache.org/jira/browse/HBASE-724
> Project: Hadoop HBase
> Issue Type: Bug
> Components: regionserver
> Affects Versions: 0.1.3
> Reporter: Sebastien Rainville
> Attachments: testdeleteall.patch
>
>
> When deleting a row, that row seems to be deleted and everything to be normal
> but some time afterwards that row somehow reappears (HTable.get() and
> scanners find it again). Looking at the log files it seems like deleted rows
> reappear after a flush of the memcache or after the compaction...
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
