We rolled back a prompt and saw staging recover, but three serverless regions were still serving different instructions. One used the code default, another read from an environment variable, and a third served a stale edge-cache value from the previous deployment. An unnoticed model revision introduced another variable, so our experiment results were misleading. We also lacked prompt lineage connecting an immutable prompt version with the region, cache state, model metadata, and request timestamp. What does a rollback actually prove unless every trace includes those fields and environment parity is verified before analysis? How are you enforcing immutable prompt versions across serverless regions, and do you fail closed when model or cache metadata is missing?
4 Answers
Put the prompt ID, model revision, region, configuration source, and cache hit or miss status on every trace. Make those fields mandatory rather than optional, and reject or flag requests when the metadata cannot be recorded.
Expose a small status endpoint in every region that reports the active immutable prompt ID, model revision, configuration source, and cache state. An automated parity check can compare those values before a rollout or rollback is considered successful.
Prompt lineage is the key gap. Every request should connect the immutable prompt version to the deployment, region, model metadata, cache state, and request time. I’d also keep an application-level version hash so rollback validation does not depend on a single dashboard or tracing system.
Immutable versions only help if deployments are forced to reference one specific version. Don’t let regions fall back to code defaults or independently resolved environment variables. If the prompt ID, model revision, or cache metadata is missing or inconsistent, fail closed or route traffic away until parity is restored.
Prompt ID and cache state would have saved us the most time. We were effectively guessing which version each region had served.

That makes sense. If the fields are optional, they will probably be missing exactly when we need them most.