On Fri, Sep 10, 2021 at 10:29:21PM +0800, "brookxu.cn" [off-list ref] wrote:

Through events and events.local, we can determine which node has
insufficient resources. For example, when the ‘events’ is large, we traverse
down and use events.local to determine which node has insufficient
resources.

IIUC, this works in situations when the limits are set in decreasing
fashion (from root down) till very (controller) leaves. That's a valid
config and you're right that following 'max' events gets you to the
misbehaving/underprovisioned cgroup.

'fail' counter does not seem to provide more effective
information in this regard. When 'fail' is big, it seems that we still need
to use events and events.local to determine the node of insufficient
resources.
I am not very sure what details can we learn through 'fail' counter.

If there's a limit on certain level with otherwise unconstrained cgroup
structure below (a valid config too), the 'fail' counter would help
determining what the affected cgroup is. Does that make sense to you?

The log messages as implemented currently, aren't as useful as proposed
'fail' counter (they would need report 'cg' path, not 'i').

I see justification for 'fail' events as a replacement for the dropped
log messages. Although it's not a complete replacement due to longer
persistence of the log, illustrated e.g. with the short-lived containers
whose cgroups/fail counts are gone).


Michal