[
https://issues.apache.org/jira/browse/SAMZA-146?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jakob Homan updated SAMZA-146:
------------------------------
Attachment: SAMZA-146.patch
Patch that adds a (capped) doubling back off to the calls to refresh, based on
the result of the previous call. This allows us to avoid the cost of polling
each needed system every time and still be responsive when new messages are
found.
In production this increased the throughput of the topic heavy jobs 4-5x and
dropped the no-results-returned-from-polling rate down to around 70%.
The maximum back off is sensitive to the number of topics the job is consuming
and is therefore made proportional to that. Jobs consuming from just a few
topics have a small backoff. Testing against TestSamzaContainerTest.html shows
improvement in jobs with large numbers of topic partitions and no impact for
ones with small numbers. This test doesn't simulate the distribution of
message delivery rate described above and seen in production. With that type
of distribution, the improvement is much more dramatic, as described above.
> Throughput degrades unreasonably as the number of topic-partitions increases
> ----------------------------------------------------------------------------
>
> Key: SAMZA-146
> URL: https://issues.apache.org/jira/browse/SAMZA-146
> Project: Samza
> Issue Type: Bug
> Components: container
> Affects Versions: 0.7.0
> Reporter: Jakob Homan
> Assignee: Jakob Homan
> Fix For: 0.7.0
>
> Attachments: SAMZA-146.patch
>
>
> Currently we poll each stream partition in need of messages before each call
> to process. For jobs with a large number of partitions, and particularly
> when those partitions are low volume, this dramatically impacts performance.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
