[
https://issues.apache.org/jira/browse/KUDU-1952?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15938909#comment-15938909
]
Todd Lipcon commented on KUDU-1952:
-----------------------------------
This also hurts load performance since the IO load for looking up blooms or
keys is concentrated on one disk. I see a factor-of-ten difference in 'await'
times from best to worst disk once a write workload entered disk-seek-bound
territory.
> round-robin block allocation can place all blocks for a given column on one
> disk
> --------------------------------------------------------------------------------
>
> Key: KUDU-1952
> URL: https://issues.apache.org/jira/browse/KUDU-1952
> Project: Kudu
> Issue Type: Bug
> Components: fs, perf
> Affects Versions: 1.3.0
> Reporter: Todd Lipcon
>
> Currently the LBM allocates across disks in round-robin order. I ran a test
> using 4 data disks and a table with 12 columns. This caused very uneven
> allocation across disks, since all data for column N ended up on disk N%4.
> This also makes single-column scans slow since they read from only a single
> disk.
> In a more mixed workload this is likely to be less problematic, but still it
> seems like we should take available space into consideration for allocation,
> and/or add more randomness.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
