Canary model
Difficulty 3 of 5
A new ranking model has completed training. Order its rollout. The held-out test, shadow check, and canary gates all pass. The policy requires each gate before the next stage and keeps the old model ready for rollback.
Put these steps in order
- Expand to all traffic and keep monitoring and rollback ready
- Finish the required canary window and compare all gates
- Send a small live group to the candidate and retain the baseline
- Run the candidate on live inputs in shadow and pass its gates
- Pass the held-out evaluation against the current model
- Define quality, error, latency, cost, and rollback gates
Goals and limits
- p99 latency
- under 200 ms
- Monthly cost
- under 10 USD/mo
- Traffic
- 100 req/s
Concepts
How players did
Player results load from the server.