Post

Today in "weird shit Kubernetes does": did you know that if you try to do an in-place resize for a pod, the result can be one of three states?

Those states are "yes, did it", "nope, sorry, can't fit", and "well it can't fit RIGHT NOW, but it might fit sometime in the far distant future, so I'm just gonna sit here forever and hope" (The technical term for the last state is "Deferred", in case you're curious)

The main way you can get into this state (aside from just "durrrr it took containerd 20 seconds longer than expected) is if you have pod requests != limits.

Hey kubelet, lemme give you a clue! The other pods aren't going to stop being greedy jerks! Like, ever!

Anyways, I guess this is just yet-another reason why you should always set requests=limits.

Genuinely the most controversial statement I've seen in my career. (Even though I subscribe to it.)