An instance of this failure is here. The documentation on Cirrus indicates that this should be a rare occurrence. The documentation also states that such tasks "will require rescheduling and automatically restart tasks". I don't think our presubmits are doing this since I waited about 30 mins for the automatic restart before manually re-running the job. If autorollers and external contributors are gated on such checks, we should consider mechanisms to detect and re-run these jobs on this edge case.