cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Eddie (JIRA)" <j...@apache.org>
Subject [jira] Commented: (CASSANDRA-2158) memtable_throughput_in_mb can not support sizes over 2.2 gigs because of an integer overflow.
Date Sun, 13 Feb 2011 17:46:57 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-2158?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12994124#comment-12994124
] 

Eddie commented on CASSANDRA-2158:
----------------------------------

I think I have a usecase where larger memtables would help a lot.  I have a combination of
fat columns that can update frequently, and I have lots of memory (currently 96 gb).  I know
I could also handle this by putting more boxes in the cluster, but I think I can get a lot
more out of the boxes I have.  I am experimenting with breaking up my cf into multiple ones
to get the same effect as the bigger sstable.  So far it seems to perform well, but feels
hacky.

Even if you decide to have a hard limit on the memtable size, it should probably fail loudly
instead of generating hundreds of sstables.  With my understanding of the current defaults,
any default install of cassandra with more than 32 gb of memory will default to this state
and will be hard for new users to understand (32/2 -> 16 gig heap | 16/8 -> 2gb default
CF memtable throughput).  I would much prefer the option than the hard limit though :).

> memtable_throughput_in_mb can not support sizes over 2.2 gigs because of an integer overflow.
> ---------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-2158
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-2158
>             Project: Cassandra
>          Issue Type: Bug
>    Affects Versions: 0.7.0
>            Reporter: Eddie
>
> If memtable_throughput_in_mb is set past 2.2 gigs, no errors are thrown.  However, as
soon as data starts being written it is almost immediately being flushed.  Several hundred
SSTables are created in minutes.  I am almost positive that the problem is that when memtable_throughput_in_mb
is being converted into bytes the result is stored in an integer, which is overflowing.
> From memtable.java:
>     private final int THRESHOLD;
>     private final int THRESHOLD_COUNT;
> ...
> this.THRESHOLD = cfs.getMemtableThroughputInMB() * 1024 * 1024;
> this.THRESHOLD_COUNT = (int) (cfs.getMemtableOperationsInMillions() * 1024 * 1024);
> NOTE:
> I also think currentThroughput also needs to be changed from an int to a long.  I'm not
sure if it is as simple as this or if this also is used in other places.

-- 
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Mime
View raw message