Observe across layers
A slow job may be constrained by the workload, storage, networking or a degraded device. Preserve metrics at the job and infrastructure levels with a common timeline. Device activity alone may not explain whether useful progress is being made.
Design recovery before scale
Checkpoint policy, failure detection, restart and rescheduling should be exercised before a critical run. Define who is on call, what evidence is collected and how responsibilities change between a cloud provider and an internal team.
Control both people and services
Resource access should reflect a person or service’s purpose. Review project boundaries, permissions, logging and removal of access. Our readiness review captures missing evidence; it does not connect to an account or verify a provider’s controls.