CORE GUIDE
Rollout and fallback
AI releases are safer when exposure increases gradually and the system has a tested fallback that preserves critical product behavior when the new model, prompt or dependency fails.
Mental model
A rollout is an experiment with bounded blast radius. Fallback is the predesigned degraded path the product can enter when evidence crosses a veto or a dependency becomes unavailable.
Why it matters
Model and prompt changes can improve global metrics while breaking a high-value slice. External providers can also degrade without notice. Shipping everything to 100% traffic turns uncertainty into a large incident. Progressive exposure and explicit fallback make evidence collection reversible and reduce the cost of discovering a bad assumption.
01
Define gates before increasing exposure
Choose rollout units such as internal traffic, small percentages, regions or customer cohorts. Define acceptance metrics, veto slices and observation windows before launch. Make fallback concrete: previous model version, rule-based behavior, human handoff or reduced capability. Test that the fallback can actually be activated and that state remains compatible across transitions.
02
Example: new model improves average quality but breaks tool calls
A model upgrade performs better on broad evaluation but produces invalid tool arguments for one enterprise workflow. A 5% rollout exposes the regression before full release. The gate blocks expansion for that slice and routes affected traffic to the previous model while engineers investigate, rather than rolling the entire product backward blindly.
Common failure modes
- Increasing traffic before the previous rollout stage has enough evidence.
- Calling a fallback 'available' without testing state and dependency compatibility.
- Using only a global quality average and missing veto-worthy regressions.
Engineering heuristics
- Pre-register rollout gates, vetoes and observation windows.
- Keep fallback paths simple enough to exercise regularly.
- Verify state and idempotency behavior when switching between versions.
Takeaways
- 01Rollout limits blast radius while learning from production.
- 02Fallback is an engineered product path, not an emergency wish.
- 03Release decisions need explicit veto slices.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.