Why rollback matters
Even a well tested model can fail in production through a data shift or a serving bug. A rollback strategy is the plan to return to a known good model fast, before damage grows.
Deployment patterns that ease rollback
- Blue green keeps the old version running while the new one takes traffic, so reverting is flipping back instantly.
- Canary routes a small slice of traffic to the new model first, limiting blast radius if it misbehaves.
- Versioned artifacts mean the previous model is one promotion away in the registry.
What makes rollback fast
- A clear trigger such as an error rate or guardrail metric crossing a threshold.
- A previous version kept warm and ready to receive traffic.
- An automated switch rather than a manual rebuild.
The principle
Rollback should be faster and lower risk than fixing forward under pressure. Design for it before you deploy, not during the incident.
Key idea
A rollback strategy keeps the previous model ready and uses patterns like blue green or canary so a failing deployment can be reverted instantly on a clear trigger.