At this point I think the gains would be much less from further splitting. I am looking into parallel build restoration. Is it true that there were primarily the two failures?
https://issues.apache.org/jira/browse/BEAM-5207 :beam-runners-apex:compileTestJava https://issues.apache.org/jira/browse/BEAM-5035 :beam-examples-java:compileTestJava The problem seems extremely likely to be our homebrew shadowTestJar. I looked at it and nothing obvious jumped out at me. But the point should be moot - dependencies like these are simply incorrect. It seems likely to be a (common) misunderstanding of what a test jar is. See https://issues.apache.org/jira/browse/BEAM-3138 and https://issues.apache.org/jira/browse/BEAM-3573. Nothing should ever have a compile dependency on a class in another module's test jar. Runtime is OK for scraping test suites like ValidatesRunner (though this too is fairly fancy and I've never seen other projects do it). So fixing our deps might make the whole problem go away. It looks like as a first step we need to promote some misc. code in the test jar to either the main GCPIO jar or a supplementary _main_ jar of a test-utils module. Kenn On Thu, Oct 25, 2018 at 4:21 PM Udi Meiri <eh...@google.com> wrote: > Perhaps you could further split pre-commits along logical lines: examples, > core, IO, runners, etc. > > > On Thu, Oct 25, 2018 at 4:02 PM Scott Wegner <sc...@apache.org> wrote: > >> Splitting into separate jobs that can be parallelized seems like a win >> for as long as Gradle task parallelization is disabled. Thanks for driving >> this improvement. >> >> > I'm in favor of (simple!) build breaks going in before precommits >> finish, on the promise that the offending test(s) passed locally. >> >> +1. Some of our post-commits run even longer (many hours); it's not >> practical to wait for a full run to merge a fix. For targeted >> fixes/rollbacks to restore Jenkins test signal, I support using a passing >> local run as a stand-in for a full Jenkins execution. Note that if you run >> gradle with the --scan flag, it will produce a Build Scan URL to submit >> with you PR. >> >> On Thu, Oct 25, 2018 at 5:39 AM Kenneth Knowles <k...@apache.org> wrote: >> >>> The split did seemingly trim about 30 minutes off the Java precommit. Of >>> course the difference between 50 and 80 minutes won't qualitatively change >>> much. I don't see any other obvious and easy wins. I still like the split >>> for the separation of signals. >>> >>> Kenn >>> >>> On Tue, Oct 23, 2018 at 2:47 PM Robert Bradshaw <rober...@google.com> >>> wrote: >>> >>>> On Tue, Oct 23, 2018 at 11:28 PM Kenneth Knowles <k...@apache.org> >>>> wrote: >>>> >>>>> Hi all, >>>>> >>>>> Java Precommit duration is about 1h15. That is quite a burden. >>>>> Especially if something gets broken. >>>>> >>>> >>>> I'm in favor of (simple!) build breaks going in before precommits >>>> finish, on the promise that the offending test(s) passed locally. Short of >>>> that, we can roll back. >>>> >>>> If it were cheap to get a fast "this is probably good" signal, that >>>> could be useful as well, though once you hit the "I'm waiting long enough >>>> to go do something else" the difference between 20 minutes and 80 minutes >>>> is not that huge. >>>> >>>> >>>>> We turned off parallel builds, which we really need to re-enable. >>>>> >>>> >>>> +1 >>>> >>>> >>>>> But beyond that, I see low-hanging fruit that would most appropriately >>>>> be a separate Jenkins job. >>>>> >>>>> Here's a scan of a successful run: >>>>> https://scans.gradle.com/s/2s4bd5hc45wuy/timeline >>>>> >>>>> * 17m :beam-runners-google-cloud-dataflow-java-examples:preCommit >>>>> * 4m >>>>> :beam-runners-google-cloud-dataflow-java-examples-streaming:preCommit >>>>> These are integration tests that should have their own job & status >>>>> anyhow. We lumped them in because Maven can't do separate tests. Gradle >>>>> makes this cheap and easy. >>>>> >>>>> Then there are these which are the only other tasks over 1m: >>>>> >>>>> * 2m :beam-runners-google-cloud-dataflow-java-legacy-worker:test >>>>> * 2m :beam-runners-google-cloud-dataflow-java-fn-api-worker:test >>>>> * 2m :beam-sdks-java-nexmark:test >>>>> * 1m :beam-sdks-java-io-google-cloud-platform:test >>>>> * 1m :beam-sdks-java-io-hbase:test >>>>> * 1m :beam-sdks-java-extensions-sql:test >>>>> >>>>> Maybe not worth messing with these. Also if we remove all the >>>>> shadowJar and shadowTestJar tasks it actually looks like it would only >>>>> save >>>>> 5 minutes, so I was hasty in thinking that would solve things. It will >>>>> make >>>>> interactive work better (going from 30s to maybe <10s for rebuilds) but >>>>> won't help that much for Jenkins. >>>>> >>>>> Kenn >>>>> >>>> >> >> -- >> >> >> >> >> Got feedback? tinyurl.com/swegner-feedback >> >
