This message was deleted.

Are you deploying in Kubernetes?

Yeah we are

Us too. We want to rely on Kubernetes+Docker’s namespace and cgroup so we use the smallest number of task slots. Two per MM.

interesting, that'd result in a lot of middlemanager nodes for a large cluster

yes, but the memory behavior is predictable. If you have too many peons in a pod, one bad subtask can bulldoze the other peons, creating difficult to track errors.

thanks for the info

we also use 2 per MM on k8s