Code Coverage runs on every PR touch, across Arm and x86, exercising unittests and handpicked JSTests. With each task uploading results to Coveralls, this amounts to ~50 "jobs" per "build" (Coveralls' terminology). This amount of data aggregation impacts Coveralls' throughput, to the point that data is slow to resolve, or timeouts occur.

Though we are paying customers, we are still confined to their "fair use" policy:

• Every customer's first 100 uploads/day / first 40 uploads/hr are processed the same way using FIFO across a set of primary background job queues.
• After you exceed one of those thresholds, though, your background jobs move to a secondary set of queues that are dequeued slightly less frequently than the main queues. Typical dequeue ratio is about 2/3 to 1/2 that of the main queues.

We are negotiating with Coveralls on better support. In the meantime, I'd like to remove the x86 variant (cutting our upload tasks by 50%) and study its effects (on our data availability and coverage metrics).