The root of this is definitely GCS. We noticed some keys in a GCS bucket go inaccessible since like 19:15-19:20ish PDT, and then we noticed increased timeouts/503's ramping up at 19:30 and plateauing at 19:45. But error rate for us seems to have gone back to nominal levels since 20:10... but the keys that originally went unavailable are still returning 503's.
Fun times.