What is meant by "system-wide level" here? Is it referring to vm.swappiness sysctl?

If that's the case, the downside I see there is it's system wide. For example maybe there are some usecases where you want certain workloads to swap often (and as such set a high swappiness), but for system workloads (e.g those in /system.slice, i.e. kubelet, container runtime, etc), you don't want them to swap at all (and want a zero swappiness). If it's a system wide setting I guess you lose the ability to make those types of adjustments.

Not sure how important that is and probably a post alpha thing, but maybe something to consider :)

This KEP doesn't enable setting per-workload swap utilization, so I think it's YAGNI on per-workload swappiness configuration in the CRI. When we're ready to add per-workload swap configuration to k8s, then we should add it IMO.

but for system workloads (e.g those in /system.slice, i.e. kubelet, container runtime, etc), you don't want them to swap at all (and want a zero swappiness).

Typically one wouldn't run the container runtime, kubelet, etc. using k8s though, they're prerequisites for a node that run directly on the host. So potentially this is still individually configurable for them, no?

When we're ready to add per-workload swap configuration to k8s, then we should add it IMO.

Make sense!

So potentially this is still individually configurable for them, no?

I guess what you would do in that case is set vm.swappiness sysctl so that it applies system wide, but then you would have to edit all of your systemd units for kubelet, container runtime etc and override their swap settings. Might be helpful to note somewhere if that's the recommended path to configure swappiness since I would imagine most existing kubelet/container runtime units don't override swap settings.

edit: actually looking at the systemd config options, I don't think it exposes overriding swap for individual units... https://www.freedesktop.org/software/systemd/man/systemd.resource-control.html

i would prefer to never introduce the MemorySwapLimit parameter.

I would update this section to say:

1. kubelet detects available swap on startup (all available swap is basically allocatable)
2. kubelet enables reservation of swap to reduce node allocatable swap (apha or beta period)

This needs a little more clarification.

The container runtime will write the swap settings to the container level cgroup. This will include ephemeral containers. We need to account for the proper setting that is applied at the pod cgroup boundary. Please list addressing this as a blocker to beta for the feature

We may want to include this as part of CRI stats as well.