Do the underlying machines that the GitHub Actions inconsistent in the specs that they have?
Given a random set of runs, we sometimes see runners be assigned to machines that have half the IOPS (and less CPU) than other machines. Here is an example:
IO CPU 4284 438 4126 394 2609 338 Stall 3870 369 4329 442 2548 339 Stall
This manifests in our test suite as the tests sometimes taking 2x as long to run.