[
https://issues.apache.org/jira/browse/CASSANDRA-15686?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17080546#comment-17080546
]
Kevin Gallardo commented on CASSANDRA-15686:
--------------------------------------------
{quote}and saw more failures than normal (normally ~1-3 failures, had 9; most
known flaky tests)
{quote}
Hm that's a bummer, I think I saw similar failures when testing out too, but
did not investigate more and thought they were just flaky. Do you think these
could be fixed, or put in another bucket to run in another non-parallel suite?
For reference, this is the tests I had run a little while ago (about 2 weeks)
for the *unit* tests only, on parallelizing the tests on current LOWRES config:
- [Run of p=4 and i=medium with no change to the current build (which means it
uses 1 runner by
default)|https://app.circleci.com/pipelines/github/newkek/cassandra/4/workflows/16897387-cbf0-49e8-82bc-327d75885256]
- [Same hardware, but hardcoding
{{-Dtest.runners=2}}|https://app.circleci.com/pipelines/github/newkek/cassandra/5/workflows/56b341b5-8f74-414e-852e-52cba8acd28a]
You can see the runtime goes from 19min to 12.
I am looking at old experiments I ran, and it seems I was able to run with 8
runners on unit tests, and all that was reported seem like failures from flaky
tests:
[run|https://app.circleci.com/pipelines/github/newkek/cassandra/22/workflows/68bf0582-2b43-4b71-a610-e1e76c6ff418/jobs/90]
and [corresponding
config|https://github.com/newkek/cassandra/blob/00103c51d3c6b1fd0c534b649e146bbe45bcdf11/.circleci/config.yml#L975]
> Improvements in circle CI default config
> ----------------------------------------
>
> Key: CASSANDRA-15686
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15686
> Project: Cassandra
> Issue Type: Bug
> Components: Build
> Reporter: Kevin Gallardo
> Priority: Normal
>
> I have been looking at and played around with the [default CircleCI
> config|https://github.com/apache/cassandra/blob/trunk/.circleci/config.yml],
> a few comments/questions regarding the following topics:
> * Python dtests do not run successfully (200-300 failures) on {{medium}}
> instances, they seem to only run with small flaky failures on {{large}}
> instances or higher
> * Python Upgrade tests:
> ** Do not seem to run without many failures on any instance types / any
> parallelism setting
> ** Do not seem to parallelize well, it seems each container is going to
> download multiple C* versions
> ** Additionally it seems the configuration is not up to date, as currently
> we get errors because {{JAVA8_HOME}} is not set
> * Unit tests do not seem to parallelize optimally, number of test runners do
> not reflect the available CPUs on the container. Ideally if # of runners == #
> of CPUs, build time is improved, on any type of instances.
> ** For instance when using the current configuration, running on medium
> instances, build will use 1 junit test runner, but 2 CPUs are available. If
> using 2 runners, the build time is reduced from 19min (at the current main
> config of parallelism=4) to 12min.
> * There are some typos in the file, some dtests say "Run Unit Tests" but
> they are JVM dtests (see
> [here|https://github.com/apache/cassandra/blob/trunk/.circleci/config.yml#L1077],
>
> [here|https://github.com/apache/cassandra/blob/trunk/.circleci/config.yml#L1386])
> So some ways to process these would be:
> * Do the Python dtests run successfully for anyone on {{medium}} instances?
> If not, would it make sense to bump them to {{large}} so that they can be run
> successfully?
> * Does anybody ever run the python upgrade tests on CircleCI and what is the
> configuration that makes it work?
> * Would it make sense to either hardcode the number of test runners in the
> unit tests with `-Dtest.runners` in the config file to reflect the number of
> CPUs on the instances, or change the build so that it is able to detect the
> appropriate number of core available automatically?
> Additionally, it seems this default config file (config.yml) is not as well
> maintained as the
> [{{config-2_1.yml}}|https://github.com/apache/cassandra/blob/trunk/.circleci/config-2_1.yml]
> (+its lowres/highres) version in the same folder (from CASSANDRA-14806).
> What is the reasoning for maintaining these 2 versions of the build? Could
> the better maintained version be used as the default? We could generate a
> lowres version of the new config-2_1.yml, and rename it {{config.yml}} so
> that it gets picked up by CircleCI automatically instead of the current
> default.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
