Repository navigation
Buildbots fail when new files are added #81224
Description
Activity
As reported here, I submitted a pull request that passed all tests locally and in CI, but when accepted, build bots started to fail as a result of new files having been added to the project. It seems it's necessary in Makefile.pre.in to duplicate the git manifest to declare those directories... but only for buildbot builds. I don't fully understand why that is the case, but it would be nicer if there were protections from this footgun, especially since this issue may manifest several times in the same PR.
I can think of some ways to prevent this undesirable behavior:
- Include buildbot builds in the GitHub merge request checks.
- Add a GitHub bot that checks merge requests that create directories and alert the review of directories created if Makefile.pre.in doesn't include changes reflecting those directories.
- Remove the reliance on a separate directory listing in Makefile.pre.in.
- added3.8 (EOL)end of lifeend of lifetestsTests in the Lib/test dirTests in the Lib/test dir
on May 25, 2019 In #57774, @yan12125 provides this reference to the buildbot code that copies the code and thus imposes the requirement to declare source directories.
The issue here is not with buildbots, but with installation on POSIX platforms. We do now have a couple of buildbots that install Python to a local location before running the tests, which is what flushes this out (see https://github.com/python/buildmaster-config/blob/master/master/custom/factories.py#L134-L147; they simply run
make installbefore running the tests).Adding these buildbots as pre-merge CI is not currently an option due to security implications (I don't want unreviewed code running on my home network). I have plans to eventually allow certain builders to be run pre-merge iff the
awaiting mergelabel is present on the PR, but I haven't had time to work on that yet.It might be possible to adjust one of the Travis builds to install before running tests, but that leaves some other tests un-run, which just relocates the problem.
Removing reliance on an explicit listing of directories sounds nice, but does open up the possibility of installing more than expected if run from a dirty checkout.
What about adding a "new directories added to Makefile.pre.in" check to Tools/scripts/patchcheck.py?
I encountered this issue again today in bpo-46118.
I started looking into creating the patchcheck, but I realize it may be a little tricky to detect the introduction of a new folder, because
git diffdoesn't actually report new folders, and my initial research indicates it's far from straightforward.I thought maybe git diff --dirstat might help, but it doesn't. It reports 100% different for a new file in a new folder and same for a new file in an existing folder.
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields: