it does seem like they've gotten flakier in the past few days. I looked at a couple of failures and it seemed like the GHA nodes are slower than they used to be, and sometimes the clock skew we add (+ the slower GHA node) makes cockroach eject the node from the cluster (which then means it can't be properly tested). if you see a failure, the artifacts for the job will include the crdb node logs; we can start collecting them to see if my thesis is right