Try to keep master tablets on separate nodes
derekperkins opened this issue · 0 comments
In most deployments, master tablets will see much higher resource consumption on cpu and network usage, and RAM to a lesser extent. This can't be handled with standard affinity/anti-affinity since reparenting happens without any impact on k8s labels. I would love for the operator to monitor how many master tablets live on each node and try to keep things balanced.
The least intrusive way would be to just reparent tablets when certain nodes have too many, though this could cause a cascading effect.
More intrusive would be to actually reschedule pods to different nodes. This would give you more control, but would involve more disk attaching/detaching, and in the case of local ssd mounts, full restores of data.