[jira] [Comment Edited] (CASSANDRA-6269) Add ability to ignore L0 on CF level

["Matt Kapilevich (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-6269?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13846789#comment-13846789 ] 

Matt Kapilevich edited comment on CASSANDRA-6269 at 12/12/13 9:54 PM:
----------------------------------------------------------------------

I would love to get 5371, but we can't upgrade to 2.0 until it's more mature, and part of DSE. Do you have objections to us committing this patch ourselves on a 1.2 branch?


was (Author: matvey14):
I would love to get 5371, but we can't upgrade to 2.0 until it's more mature, and part of DSE. Do you have objections to us committing this fix ourselves on a 1.2 branch?

> Add ability to ignore L0 on CF level
> ------------------------------------
>
>                 Key: CASSANDRA-6269
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-6269
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Matt Kapilevich
>         Attachments: L0-vs-availability.png
>
>
> One of our CF's is written to only from a batch process. We use Cassandra's bulk-load utility to load the data. When the load happens, the number of tables in L0 increases, and then comes back down as they are compacted. While the number of tables in L0 is high, there's increased load on the node, and read availability suffers, since L0 is unsorted, and therefore lookups against L0 are inefficient.
> This all works-as-designed, and issues around L0 are known.
> I think it would be a great addition to disable reading from L0, settable on CF-level, as one of Leveled Compaction options. In our case, because the data is written by a batch process, we are fine waiting a little longer while L0 is compacted away. However, the decrease in availability rate while this is happening is an issue for us.
> I would propose to add "disable_reads_from_L0" parameter to compaction_strategy_options, with default being false. In cases when availability is much more important than consistency, like ours, user can set it to true.
> I've attached a graph that shows the relationship between our availability rate and number of tables in L0.



--
This message was sent by Atlassian JIRA
(v6.1.4#6159)