[ 
https://issues.apache.org/jira/browse/CASSANDRA-10505?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14955275#comment-14955275
 ] 

Brandon Williams commented on CASSANDRA-10505:
----------------------------------------------

So it occurs to me that maybe we only have one problem here, because there is a 
way to get duplicates I didn't originally consider:

Nodes X, Y RF=2.  Client does a write containing a RT to X at ONE.  X 
determines the local deletion time, but times out replicating to Y.  However, 
the write actually does succeed on Y, X just doesn't know that.  Now X has a 
hint for for Y for that write, and 10 minutes later delivers it when the HH 
timer fires.  In the mean time, Y has flushed the original write to an sstable. 
 When Y receives the hinted write, this too is eventually flushed to an 
sstable.  Now we have the exact same write in two sstables, and we know 
compaction will do the rest and amplify this repeatedly the more it happens to 
compact.

> Duplicate range tombstones
> --------------------------
>
>                 Key: CASSANDRA-10505
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-10505
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>            Reporter: Brandon Williams
>             Fix For: 2.1.x
>
>
> I don't have a way to reproduce this just yet, but I have the sstables to 
> prove it and can provide those offline.  What I can reproduce, however, is 
> the magnification of this problem via compaction.  Here's what I know so far:
> * under some condition, we flush the exact same RT more than once in the same 
> sstable.  Same timestamps, everything.
> * Once this happens in multiple sstables, compacting them causes the 
> duplication to grow.  The more this occurs, the worse the problem gets.
> This continues until gc_grace has passed, then they are all evicted normally. 
>  I'd at least like to take a look at the compaction amplification for now, 
> that way if the condition does occur, at least compaction can help it, rather 
> than exacerbate it.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to