On Wed, Jul 17, 2019 at 10:22 AM Wes McKinney <wesmck...@gmail.com> wrote: > > hi folks -- I noticed this last night on > https://github.com/apache/arrow/pull/4841 and it surprised me. Others > may not be aware. > > We have been using builds on Appveyor and Travis CI to decide whether > to merge PRs. The trouble is these builds are not equivalent to the > builds that Travis runs inside the PR (using the apache/arrow build > queue). The differences are: >
*missing crucial detail: "builds on personal forks" > * They do not take into account changes in master (IOW to test if the > build works after `git merge`) > * They only test the latest commit versus the previous one in the branch > > This latter item is insidious, because of the `detect-changes.py` > script. Suppose that you have a large PR that touches many components, > and you push a commit that only affects one of them. Then the > detect-changes.py script will cause Travis to only run builds for the > affected component in the most recent commit. > > Here's an example of such a spurious build > > https://travis-ci.org/wesm/arrow/builds/559745190 > > There are a few ways we can mitigate this last issue: > > * If you need a faster build, you can squash your commits and rebase > on master before pushing to make sure that Travis "sees" everything. > Note this still carries risk of conflicting changes in master causing > a broken build post-merge > > * We can change the Travis configuration to try to detect whether or > not we are testing a PR -- the detect-changes.py logic is really only > intended to speed up builds in apache/arrow > > Overall, I think we need to accelerate our exodus from Travis CI since > it's hurting the project's productivity to be waiting so long on > builds. We've moved a couple of jobs to be Docker-based but we have > quite a lot more work to do to decouple ourselves. > > Thanks > Wes
