[
https://issues.apache.org/jira/browse/BEAM-7058?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16832858#comment-16832858
]
Ankur Goenka commented on BEAM-7058:
------------------------------------
After some debugging and discussion with Alex, here is the summary.
* Python statesampler_slow is not implemented
* Python statesampler_fast requires cython to be installed so the sampling
only works with cython
* statesampler_fast wakes up and checks the state every so often and based on
the current state, increases associated time. The statesampler does not query
the system clock and just add fixed wait time to the corresponding state. This
is done to avoid system calls and improve performance.
* Its expected that over a large run, statesampler_fast will converge more
consistent values but for small pipelines, it will be very unpredictable.
* As in the test we are only having 100 elements, it is very likely that we
don't correctly compute the time related metrics.
> Python SDK metric process_bundle_msecs reported as zero
> -------------------------------------------------------
>
> Key: BEAM-7058
> URL: https://issues.apache.org/jira/browse/BEAM-7058
> Project: Beam
> Issue Type: Bug
> Components: runner-flink, sdk-py-harness
> Reporter: Thomas Weise
> Assignee: Alex Amato
> Priority: Major
> Labels: metrics, portability-flink, portable-metrics-bugs
> Attachments: test-metrics.txt
>
>
> With the portable Flink runner, the metric is reported as 0, while the count
> metric works as expected.
> [https://lists.apache.org/thread.html/25eec8104bda6e4c71cc6c5e9864c335833c3ae2afe225d372479f30@%3Cdev.beam.apache.org%3E]
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
