Safe Model Rollout and Feature-Cost Reduction for ML Serving

System Design · Disney · Medium

This discussion covers two separate scenarios. In the first, a newly built inference model has to be proven out before it is allowed into production. Explain how you would roll it out, how you would design an A/B test to evaluate it, which metrics you would track, and how you would measure the lift it delivers. In the second, a different model that consumes more than 2000 features has become expensive to run, and the goal is to bring its operating cost down. Address how you…

Loading the rest of the question…