A useful tool for investigating test flakiness is my Jenkins Test Explorer service, running at https://spark-tests.appspot.com/
This has some useful timeline views for debugging flaky builds. For instance, at https://spark-tests.appspot.com/jobs/spark-master-test-maven-hadoop-2.6 (may be slow to load) you can see this chart: https://i.imgur.com/j8LV3pX.png. Here, each column represents a test run and each row represents a test which failed at least once over the displayed time period. In that linked example screenshot you'll notice that a few columns have grey squares indicating that tests were skipped but lack any red squares to indicate test failures. This usually indicates that the build failed due to a problem other than an individual test failure. For example, I clicked into one of those builds and found that one test suite failed in test setup because the previous suite had not properly cleaned up its SparkContext (I'll file a JIRA for this). You can click through the interface to drill down to reports on individual builds, tests, suites, etc. As an example of an individual test's detail page, https://spark-tests.appspot.com/test-details?suite_name=org.apache.spark.rdd.LocalCheckpointSuite&test_name=missing+checkpoint+block+fails+with+informative+message shows the patterns of flakiness in a streaming checkpoint test. Finally, there's an experimental "interesting new test failures" report which tries to surface tests which have started failing very recently: https://spark-tests.appspot.com/failed-tests/new. Specifically, entries in this feed are test failures which a) occurred in the last week, b) were not part of a build which had 20 or more failed tests, and c) were not observed to fail in during the previous week (i.e. no failures from [2 weeks ago, 1 week ago)), and d) which represent the first time that the test failed this week (i.e. a test case will appear at most once in the results list). I've also exposed this as an RSS feed at https://spark-tests.appspot.com/rss/failed-tests/new. On Wed, Feb 15, 2017 at 12:51 PM Saikat Kanjilal <sxk1...@hotmail.com> wrote: I would recommend we just open JIRA's for unit tests based on module (core/ml/sql etc) and we fix this one module at a time, this at least keeps the number of unit tests needing fixing down to a manageable number. ------------------------------ *From:* Armin Braun <m...@obrown.io> *Sent:* Wednesday, February 15, 2017 12:48 PM *To:* Saikat Kanjilal *Cc:* Kay Ousterhout; dev@spark.apache.org *Subject:* Re: File JIRAs for all flaky test failures I think one thing that is contributing to this a lot too is the general issue of the tests taking up a lot of file descriptors (10k+ if I run them on a standard Debian machine). There are a few suits that contribute to this in particular like `org.apache.spark.ExecutorAllocationManagerSuite` which, like a few others, appears to consume a lot of fds. Wouldn't it make sense to open JIRAs about those and actively try to reduce the resource consumption of these tests? Seems to me these can cause a lot of unpredictable behavior (making the reason for flaky tests hard to identify especially when there's timeouts etc. involved) + they make it prohibitively expensive for many to test locally imo. On Wed, Feb 15, 2017 at 9:24 PM, Saikat Kanjilal <sxk1...@hotmail.com> wrote: I was working on something to address this a while ago https://issues.apache.org/jira/browse/SPARK-9487 but the difficulty in testing locally made things a lot more complicated to fix for each of the unit tests, should we resurface this JIRA again, I would whole heartedly agree with the flakiness assessment of the unit tests. [SPARK-9487] Use the same num. worker threads in Scala ... <https://issues.apache.org/jira/browse/SPARK-9487> issues.apache.org In Python we use `local[4]` for unit tests, while in Scala/Java we use `local[2]` and `local` for some unit tests in SQL, MLLib, and other components. If the ... ------------------------------ *From:* Kay Ousterhout <kayousterh...@gmail.com> *Sent:* Wednesday, February 15, 2017 12:10 PM *To:* dev@spark.apache.org *Subject:* File JIRAs for all flaky test failures Hi all, I've noticed the Spark tests getting increasingly flaky -- it seems more common than not now that the tests need to be re-run at least once on PRs before they pass. This is both annoying and problematic because it makes it harder to tell when a PR is introducing new flakiness. To try to clean this up, I'd propose filing a JIRA *every time* Jenkins fails on a PR (for a reason unrelated to the PR). Just provide a quick description of the failure -- e.g., "Flaky test: DagSchedulerSuite" or "Tests failed because 250m timeout expired", a link to the failed build, and include the "Tests" component. If there's already a JIRA for the issue, just comment with a link to the latest failure. I know folks don't always have time to track down why a test failed, but this it at least helpful to someone else who, later on, is trying to diagnose when the issue started to find the problematic code / test. If this seems like too high overhead, feel free to suggest alternative ways to make the tests less flaky! -Kay
