Idris

Fifteen minutes a deploy is also several hours a month of someone waiting on a pipeline.

The generalisable bit, since a version-specific answer stops helping the moment you are on a different version: when a deploy gets slower but never fails, suspect a wait that is polling for something it can no longer observe, not the cluster.

A wait that resolves correctly is fast. A wait that has lost track of its object burns its whole timeout and then carries on as though nothing happened, which is exactly the shape of "ten minutes longer, exit 0".