@nitkon

• We inadvertently lost GOOS=darwin runs on PRs and branches with the new config (Fixed by Fix travis matrix #3493)
• It looks like (disk?) perf. gets exceptionally bad when two runs happen simultaneously through similar code (IBM team reaching out via TravisCI contacts)

For the second case, the merge to master of #3272 has 2 variants of the matrix running through the make root-coverage at the same time, which is mainly snapshotter tests. For one of those runs (https://travis-ci.org/containerd/containerd/jobs/568051828) the snapshotter/btrfs test run went from 38.768s to 399.944s, and if that continues through the other snapshotters, the Travis CI timeout may be reached before tests complete.

There is also an image error which may be transient, but might need looked at as well in the "merge to master" Travis CI run: https://travis-ci.org/containerd/containerd/jobs/568051827 (transient failure happened across architectures)