Misspoke slightly here - Travis doesn't build on every commit, rather it builds on every push. So if you push multiple commits at the same time you'll only get one build.
Travis has always been doing this 2-builds business for PR's, but they used to be sending notifications to the same status name "continuous-integration/travis-ci," and github would only show the most recent message that was sent to the same status endpoint. Now the branch build sends to "continuous-integration/travis-ci/push," the PR merge build sends to "continuous-integration/travis-ci/pr," and we see both statuses on PR's that come from branches within the JuliaLang/julia repo. For PR's that come from personal forks such as https://github.com/JuliaLang/julia/pull/10250 (which was just merged, click the red X next to the commit message) we only get the "continuous-integration/travis-ci/pr" status. On Thursday, February 19, 2015 at 6:47:41 PM UTC-8, Tony Kelman wrote: > > As far as I can tell this is just a change in how Travis sends > notifications to the GitHub status API. For contributors who have commit > access to the main Julia repository and work on a pull request from a > branch there, Travis builds every commit in every branch (unless the commit > message contains "[ci skip]"). Then when a pull request is opened, Travis > does a separate build of the merge commit that would result from merging > the PR branch into the state of master at the time the build begins. > > If there's a difference in results, I suspect it's mostly down to the fact > that Julia's CI tests still have a large number of intermittent failure > modes that we haven't gotten to the bottom of yet. In some cases the > failures might be caused by the actual difference in the code that gets > built, the branch build runs on the state of the branch which may not > contain newer changes that have been committed to master. > > > On Thursday, February 19, 2015 at 12:19:35 PM UTC-8, Seth wrote: >> >> I noticed recently that two travis tests (each against 0.3 and nightly, >> so a total of 4 builds?) are occurring for each PR. In many cases one fails >> while the other succeeds. >> >> What's the difference between the two (sets of) tests, and when/why did >> this change happen? >> >> Thanks! >> >
