I find that somewhat often, the K8S PR builders will fail on a PR: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder-K8s/
... when the PR seems totally unrelated to K8S. I've kind of learned to ignore them in that case but that seems wrong. Are they just kind of flaky? am I imagining things? Just trying to figure out how much they're 'accurate' in catching real vs false failures.