ImageMaximumGCAge documentation recommendations about kubernetes

[haircommander]

when concieving of the feature I had the idea of something intermittent but not too infrequent: something like 2 weeks. We've yet to make official recommendations as I think we need more testing time to find the frequency that seems good.

[Joseph-Goergen]

/sig-node

[Joseph-Goergen]

/sig node

[kannon92]

/cc @haircommander @sohankunkerkar

/triage accepted
/priority important-longterm

[kannon92]

Beta feedback for kubernetes/enhancements#4210

[SergeyKanzhelev]

@haircommander how big of a disk usage spike will happen on cleanup?

Is 2 weeks coming from the idea that the autoscaler may decide to use the node again? If there are no autoscaler, is there any issues with 1 hour?

[SergeyKanzhelev]

btw, one idea for improvement of this logic is to add some grace period to wait if there are download is hapenning at this moment. So if there is an image pull ongoing when GC want's to clean up those images, we can wait a bit to minimize disk churn

[haircommander]

I think 1 hour is pretty low TBH

autoscaler may decide to use the node again

which autoscaler?

So if there is an image pull ongoing when GC want's to clean up those images, we can wait a bit to minimize disk churn

no matter the grace period we give, we'll hit this issue theoretically. Even if no container is using an image, we could hit situations where an unused image for 1 hour will be created immediately after.

My thought with the relatively high value is to increase the liklihood that an image is not just unused, but has been replaced by a newer version of the same image. The former case (unused for a bit, but will be used later) is an unfortunate side effect, and we'd use the new metric to track when the image was gc'd for reason age