[jatone]

or you know; just pace the GC mark and sweep algorithm. which is what go is doing now.

[verdagon]

Correct me if I'm wrong, but IIRC pacing would still cause a latency spike, it would just be a more strategically-timed latency spike.

[jatone]

sure - https://github.com/golang/go/issues/44167. you'll see the new design the CPU util only increases in GC CPU utilizations when you're actually allocating heavily. which makes sense you're doing more work. this should completely resolve the problem discord had; since their system was in a steady state.

[abeltensor]

You are always going to have some kind of latency spike with a sweeping GC; even if that spike is tiny.

[jatone]

not really. if you're in a steady state (like discord was) you wouldn't see any spikes. you'd have a consistent utilization. if you start allocating heavily then you would potentially see an increase. which makes sense, you're increasing your workload, utilization needs to increase. but still not necessarily a 'spike'