Travis Downs

> That's a valid point. I never encountered such a long delays with `llvm-addr2line` during a single address decoding. I only saw it that slow with `binutils`. A single address...

> The workaround of using `-isystem` is not usable because it has other impacts such as disabling warnings in those headers. This is a critical point. Essentially there is no...

It's worth noting there is a reproducer now, for at least one type of crash, see https://github.com/scylladb/seastar/issues/2697 and https://github.com/scylladb/seastar/pull/2714. So perhaps it is worth revisiting this PR as part of...

It seems like maybe what you want is to have a larger max-delta, like 100 ms you had before, but also cap the effect this can have for newly waking...

One thing I didn't understand though: did the "statement" class actually want more IO? I.e., it had a very low rate of raw consumption, i.e., low IO rate, but if...

> The problem is that we have to penalize idle groups. If we don't, they will accrue credit and then dominate the resource. If we penalize too much, they suffer...

> Now, this is what I see when run on master: OK, but in all three cases the "statement" class has exactly the same throughput, only the latency varies (that's...

> * sometimes "statement" class suffers from large latency > * there's no apparent reason for that > * one of the sources can be this wake-from-idle thing OK sure,...

> It boils down to: we want to reward statement for sleeping, but not reward compaction for sleeping. That's at least part of the problem, but isn't another part that...

I have a ~fox~ fix for this. https://github.com/scylladb/seastar/pull/2390