That doesn't contradict my conjecture that the issue was improper sharding. The jump in duration of cancellation detection from 2ms to 20ms could have been beca

That doesn't contradict my conjecture that the issue was improper sharding. The jump in duration of cancellation detection from 2ms to 20ms could have been because they were running that calculation on a single machine. Although... Ethernet latency would probably make it tough to stick to 2ms.