> On July 13, 2014, 11:45 p.m., Benjamin Hindman wrote: > > src/tests/health_check_tests.cpp, line 159 > > <https://reviews.apache.org/r/23443/diff/1/?file=629112#file629112line159> > > > > I don't understand how changing to ASSERT from EXPECT would change the > > flakiness.
Well currently the tests keeps moving on even when it fails expectation, so just having a short circuit will try to do cleanup earlier. I never was able to repro what the tests machine was seeing where it hangs forever, but I do found issues doing a clean clone and test run so I'm trying to fix those and see what's next. > On July 13, 2014, 11:45 p.m., Benjamin Hindman wrote: > > src/launcher/executor.cpp, lines 601-605 > > <https://reviews.apache.org/r/23443/diff/1/?file=629111#file629111line601> > > > > This is a hack which assumes autotools layout. At the very least we > > should have a comment here that explains what is being done. ;-) > > > > That being said, I remember at one point you had an environment > > variable that I told you to remove. That would have been much more generic > > then this, my apologies. > > > > My concern with the environment variable was that it wasn't very > > explicit and therefore discoverable. That is, if someone else launched the > > mesos-executor they wouldn't know they couldn't pass a task with a health > > check unless you set the environment variable (and they'd have to look in > > the code to figure that out)! > > > > So it seems like we should always use dirname(argv[0]), which should > > work for the installed case. But also provide the environment variable to > > override dirname[argv[0]). Even better than an environment variable would > > be a flag, but I think that might be more complicated to take on. Given > > that this is an extension of 'launcher_dir' from the slave, how about > > naming that environment variable (or flag) MESOS_LAUNCHER_DIR as well? > > > > Just a little different than what you originally had, sorry for my lack > > of foresight in the earlier comments, thanks Tim. Sounds good, let me do this change. - Timothy ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/23443/#review47701 ----------------------------------------------------------- On July 13, 2014, 10:22 p.m., Timothy Chen wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/23443/ > ----------------------------------------------------------- > > (Updated July 13, 2014, 10:22 p.m.) > > > Review request for mesos, Adam B, Ben Mahler, Jie Yu, and Niklas Nielsen. > > > Bugs: MESOS-1533 > https://issues.apache.org/jira/browse/MESOS-1533 > > > Repository: mesos-git > > > Description > ------- > > Currently the health check tests are disabled as it was flaky. Part of the > reason is that all assertions was EXPECTs instead of ASSERTs so it will > continue to execute. > Another issue was it's currently discovering the path to launch health check > executable from the command executor main argv[0] path. > However, the correct path is to launch the generated script that does library > path resolution, which is the parent of .libs. > > Also changed the command executor to also include the health state when it > kills the task due to unhealthy checks. > > > Diffs > ----- > > src/launcher/executor.cpp 9c80848 > src/tests/health_check_tests.cpp 44711fd > > Diff: https://reviews.apache.org/r/23443/diff/ > > > Testing > ------- > > make check > > > Thanks, > > Timothy Chen > >
