On Thu, Feb 07, 2019 at 04:41:57PM +0100, Johannes Schindelin wrote: > > I think this can be limited to the tests that failed, which makes things > > much faster. I.e., we run the tests at the tip of topic X and see that > > t1234 fails. We then go back to the fork point and we just need to run > > t1234 again. If it succeeds, then we blame X for the failure. If it > > fails, then we consider it a false positive. > > If you mean merge bases by fork points, I wrote an Azure Pipeline to do > that (so that I could use the cloud as kind of a fast computer), but that > was still too slow. > > Even when there are even only as much as 12 merge bases to test (which is > the current number of merge bases between `next` and `pu`), a build takes > roughly 6 minutes on Windows, and many tests take 1 minute or more to run > (offenders like t7003 and t7610 take over 400 seconds, i.e. roughly 6 > minutes), we are talking about roughly 1.5h *just* to test the merge > bases. I was assuming you're testing individual topics from gitster/git here (which admittedly is more CPU in total than just the integration branches, but it at least parallelizes well). So with that assumption, I was thinking that you'd just look for the merge-base of HEAD and master, which should give you a single point for most topics. For inter-twined topics there may be more merge bases, but I actually think for our purposes here, just testing the most recent one is probably OK. I.e., we're just trying to have a vague sense of whether the test failure is due to new commits or old. I think Junio's suggestion to just pick some common release points would work OK in practice, too. It's possible that some other topic made it to master with a breakage, but in most cases, I think these sorts of failures are often more coarsely-grained (especially if Junio pays attention to the CI results before merging). -Peff
