Comparison
Canary deploymentvsProgressive delivery
Canary deployment
the new prompt goes to two percent of traffic first, and you watch quality and cost before letting the rest through.
Rolling a prompt or model change to a small traffic slice first, watching quality and cost metrics, then widening. Treat prompts as deployable artefacts with the same ceremony as code, because they break things the same way.
Full entry →Progressive delivery
the release goes to one percent, then five, then twenty-five, and each step is only taken if the metrics stayed flat.
Releasing to progressively larger portions of traffic with automated analysis deciding whether to continue, pause or revert. It is canarying with the judgement automated and formalised into steps. It requires two things most teams lack before they start: routing that can split traffic by percentage, and a metric trustworthy enough to abort a release on.
Full entry →