accumulo-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Bill Havanki (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (ACCUMULO-2488) Concurrent randomwalk balance check needs refinement
Date Mon, 17 Mar 2014 21:26:47 GMT

    [ https://issues.apache.org/jira/browse/ACCUMULO-2488?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13938409#comment-13938409
] 

Bill Havanki commented on ACCUMULO-2488:
----------------------------------------

Thanks, that's good information. There are actually five tables, as it turns out.

Now, the question is, what should this balance check test for? Should it test that the default
balancer is working as expected? Or should it test that the outcome fits some general definition
of "balanced"? For example, that no server should be more than _n_ standard deviations away
from the mean.

> Concurrent randomwalk balance check needs refinement
> ----------------------------------------------------
>
>                 Key: ACCUMULO-2488
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-2488
>             Project: Accumulo
>          Issue Type: Test
>          Components: test
>    Affects Versions: 1.4.4
>            Reporter: Bill Havanki
>            Assignee: Bill Havanki
>            Priority: Minor
>              Labels: randomwalk, test
>
> The check for balanced tablets in the randomwalk Concurrent test too easily fails.
> Here is a real-life example from the test for the number of tablets across five tablet
servers: 2, 5, 2, 2, 3. (An old unrelated table plays into these totals.) This produces a
mean of 2.8. The cluster is considered unbalanced by the test when any server's count differs
from the mean by the larger of 1 or the mean divided by 5. In this case, 2.8/5 is less than
1, so the second tablet server fails since it has more than 3.8 tablets. Even a 4 would fail.
> Part of the problem in this particular case is that there are so few tablets, and so
few tablet servers. The cluster also seems happy to leave these counts as is, as I continue
to check it, so the test's definition of unbalanced is too narrow.
> The test needs to be refined to detect unbalanced conditions with a statistically decent
calculation.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message