Paul Gevers <[email protected]> writes: > I looked at the results of the autopkgtest of your package because it > showed up as a regression for glibc. I noticed that it regularly most of > the time [1] on amd64, while other architectures seem to be less > impacted. I notice that 2/3 tests are only run on amd64 (may I ask why), > so that might explain that. By the way "Architecture: any" is the same > as not having the field.
Thanks for filing this bug. I've been puzzled by this for quite a while, and have long suspected that this has to do with the goroutine leak detection. The reason why only amd64 is because the tests require full virtualization and networking setup. I've been analyzing over half a dozen of failed logs, and all of them failed quite mysteriously. However, while individual tests pass (--- PASS:), in each of them the suite fails immediately after TestBeforeTestWithoutLeakDetection completes. This test is designed to check for leaks from previous tests. In the latest upload, I've decided to disable the goroutine leak detection completely. I strongly suspect we are seeing that in the package build because because of the way how we are invoking the test suite. I'll keep an eye on etcd autopkgtest runs, but do feel free to reach out to me either on irc/matrix, or via another bug report and I'll have another look. Best, -rt
