Trigger

• Monitor memory PSI information via cgroups
• We can tune which PSI is more important:
  • global cgroup
  • or high-priority cgroups: init/podruntime/runtime (?)
• Threshold: which value of PSI triggers OOM

Pick target

• First, sort by high-level group:
  • kubepods (workloads)
  • podruntime (CRI, kubelet, etcd)
  • runtime (core containerd, system services)
  • init
• Second, inside kubepods we have QoS groups:
  • first priority: BestEffort
  • second: Burstable
  • last: Guaranteed
• Third, look into other attributes, e.g. OOM score.
• Fourth, look into memory max - memory current (if memory max is set).

Pick a target(s) to kill

Kill

Kill the whole cgroup.

Cooldown

Wait some time before, next kill, go to the beginning monitoring PSI.