Travis Downs

[1] Even calling it CPU utilization isn't strictly correct as it's a wall-time measurement, not a CPU time measurement (e.g., via `rusage`), so steal time will inflate reactor utilization, unlike...

I suppose also "non-polling" is a bit imprecise, it should really be "non-_busy_-polling" as we do include in reactor utilization the polls we do when not idle, i.e., while there...

> A suggestion -- the proposed clarification is out-out-ish, it says that "this is CPU usage _excluding_ what reactor does", but it could as well be opt-in-ish by saying smth...

@xemul : > How about this then: > > cpy_busy_time: Total cpu busy time in milliseconds excluding idle-polling > utilization: Average CPU busy time for the last 5-seconds I really...

@avikivity wrote: > The io context size should be sized exactly right. Maybe there is an off-by-one somewhere. I don't think that's right, at least in the sense of `--max-networking-io-control-blocks=N`...

@avikivity - here's a more concrete repro to back the claim that N iocbs is not enough to support N sockets (nor N-C sockets for small C). https://github.com/travisdowns/seastar/tree/td-eagain-repro Run the...

The symptom is that 1 (or a few) producers seem to hang for 5 minutes while producing a message, then produce a few more messages then hang again for 5...

/ci-repeat 1

PT does not handle cross-test duplicates well, so we will have to leave this open to be close at the same time as #17504.

@ztlpn - I'm looking at the behavior of `partition_balancer_planner::request_context::for_each_partition_random_order` where this crash occurs. As I understand it, any time this code suspends, it replies on `topics_map_revision` changing so that it...