Model serving design. Online (low-latency) vs batch vs streaming, A/B + shadow + canary deployment, multi-armed bandits where they fit, model-registry promotion policy, rollback. Online serving consumes resilience-patterns (timeouts, fallbacks, graceful degradation) — a slow model is worse than a missing model. Per the Resolved Decision, ships with refs for SageMaker / Vertex AI / Azure ML / KServe / Triton / BentoML / Ray Serve plus a mode-axis ref for batch vs online vs streaming. ML/AI Engineer owns; Platform/SRE operates infrastructure; deploy-release coordinates production cutover.
