How does shadow evaluation work?
mediumAnswer
- New model runs on production requests in parallel with current model.
- Predictions logged, not served.
- Compute: (1) agreement rate with production model.
- (2) predicted-metric distribution vs production.
- (3) latency + memory under real load.
- (4) if labels available (delayed), online accuracy.
- Purpose: safety check before A/B — catch latency issues, prediction anomalies, correlations with production.
- No user impact.
- Typical duration: 1-7 days.
- Standard pre-A/B step.
Check yourself — multiple choice
- Random
- New model runs on prod requests parallel to current, predictions logged not served; compute agreement + distribution + latency/mem under load + delayed accuracy; safety check pre-A/B; no user impact; 1-7 days
- Same as A/B
- Not real
Shadow eval: parallel run without serving; safety check pre-A/B.
#deployment#monitoring
Practise MLOps & Data Quality
215 interview questions in this topic.
Related questions
- How do you structure a progressive rollout with automated guardrails?
- Why do offline improvements often not translate online?
- What is Canary Analysis (Kayenta / Flagger)?
- How do you respond to a production ML incident?
- Scheduled retraining or triggered retraining?
- What is data drift and how do you detect it?