Shadow testing and gradual rollout
Infrastructure that tests a new model without touching customers.
A new model is never applied straight to traffic: it first produces decisions in shadow, then rolls out at the percentage you set in the panel; making it permanent is a separate, manual decision. Monitoring alerts but doesn't roll back; the rollback decision rests with a human.
Model card
- Application
- Shadow first
- Rollout
- By percentage
- Rollback
- Learning preserved
- Decision
- With a human
What does it do?
Shadow decisions
The new model produces decisions, but they aren't applied.
Percentage rollout
No payment is affected until you switch it on.
Enough data
No result is declared before enough samples have accumulated.
Continuous monitoring
Approval, latency and cost are monitored regularly.
Alerts, no automatic rollback
An alert fires when a threshold is crossed; the rollback decision rests with a human.
State preserved
Nothing learned is lost when reverting to the previous version.
How does it work, step by step?
- 1
Stage 1
Candidate
The new model is prepared.
- 2
Stage 2
Shadow
Decisions are produced, not applied.
- 3
Stage 3
Percentage
It rolls out to the traffic share you choose.
- 4
Stage 4
Monitoring
Results are compared.
- 5
Stage 5
Permanent decision
A human decides whether to make it permanent.
Meet Shadow testing and gradual rollout.
Let's review with our team how the model would work in your business.