3 ms·
We use https://github.com/uber-go/automaxprocs https://github.com/uber-go/automaxprocs after we joyfully discovered that Go assumed we had the entire cluster's
by sethammons 3y ago
We use https://github.com/uber-go/automaxprocs https://github.com/uber-go/automaxprocs after we joyfully discovered that Go assumed we had the entire cluster's cpu count on any particular pod. Made for some very strange performance characteristics in scheduling goroutines.
- jeffbee 3y agoMy opinion is that setting GOMAXPROCS that way is a quite poor idea. It tends to strand resources that could have been used to handle a stochastic burst of requests, which with a capped GOMAXPROCS will be converted directly into latency. I can think of no good reason why GOMAXPROCS needs to be 2 just because you expect the long-term CPU rate to be 2. That long-term quota is an artifact of capacity planning, while GOMAXPROCS is an artifact of process architecture.
- sethammons 3y agoHow do you suggest handling that?