Overload Manager should monitor epoll loop duration and polling delay as resources

[akonradi]

The Overload Manager has a configurable set of resource monitors that can be used to trigger actions when Envoy is overloaded. Monitoring the epoll loop duration and polling delay would allow proxies to take action when they are over-taxed in terms of CPU, possibly with a large number of pending events to handle. This could be done by polling the epoll stat metric, but I'm not sure that's the most efficient option.

[antoniovicente]

Would it make more sense to monitor CPU usage which should be a smoother measure of load?

Are there specific overload manager actions that make sense when under CPU pressure?

[fcantournet]

Would it make more sense to monitor CPU usage which should be a smoother measure of load?

isn't that harder to do though ? From within envoy.

Are there specific overload manager actions that make sense when under CPU pressure?

While not specific, I think stop_accepting_requests and reject_incoming_connections are both valid actions to take under load.
We've been hit by this not long ago, CPU overload leads to kubernetes liveness probe failures, leads to crashloops, leads to more overload.