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-2608) Create a randomwalk replay meta-node
Date Mon, 21 Jul 2014 15:34:40 GMT

    [ https://issues.apache.org/jira/browse/ACCUMULO-2608?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14068637#comment-14068637
] 

Bill Havanki commented on ACCUMULO-2608:
----------------------------------------

All good thoughts, thanks for them. At least for this go-round, I'd be happy with just something
minimal, so addressing some of this points may need to wait. In the meantime, some thoughts
I have from working on this:

* A recorded run takes the form of a straight-line graph. If a node in the original graph
is visited 100 times, it is represented 100 times in the recorded graph, each time with a
different ID.
* There may be some cases where the recorded test run does not have enough information to
reproduce a bug. Keith's example of the shard test generating random documents is probably
one of them. In this first attempt, I'm (so far) recording not only each node that is visited
but also the state after the visit. So, it's conceivable that a test could be updated to store
information in state (like a random seed) that would enable a replay to run in the same way.
This might not work for some tests, of course.
* Rather than attempting the complex-yet-cool task of pruning cycles from a graph, it would
be easier to checkpoint enough information in the recording, like state, so that a run could
be kicked off from some point just before a bug expresses itself, and the run would start
with the same conditions as the original run at that point. It could be an extension to the
current XML.
* The recording mechanism definitely has to cope with memory limitations. A more naïve approach
could toss out old the oldest part of the run, while something more sophisticated could periodically
dump the recording to disk ... like a tablet. :)

> Create a randomwalk replay meta-node
> ------------------------------------
>
>                 Key: ACCUMULO-2608
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-2608
>             Project: Accumulo
>          Issue Type: New Feature
>          Components: test
>            Reporter: Mike Drob
>            Assignee: Bill Havanki
>            Priority: Trivial
>             Fix For: 1.5.2, 1.6.1, 1.7.0
>
>
> It would be useful to have the ability to replay randomwalk sequences.
> If a randomwalk fails, it is often difficult to reproduce due to the random nature of
things. If current test runs were modified to output a specially crafted module representing
their sequence, then future runs could accept it as the module to run and improve reproduction
capabilities.



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

Mime
View raw message