Design ML deployment with GitHub and Jenkins
Company: Capital One
Role: Data Scientist
Category: Machine Learning
Difficulty: Medium
Interview Round: Onsite
Design an end‑to‑end ML deployment for a prediction model using GitHub and Jenkins: 1) Propose a repo layout (src/, features/, data_contracts/, tests/, notebooks/, ci/, docker/) and how you pin environments (lockfiles, Docker, CUDA matrix). 2) Explain model/data versioning (semantic model versions, commit SHAs, DVC/LakeFS, artifact store), model registry, and provenance capture so any prediction is reproducible. 3) Define CI checks (unit tests, data‑schema tests, leakage guards, determinism checks, small smoke‑train) and CD gates (minimum offline metrics, bias metrics, size/latency budgets). 4) Describe Jenkins pipeline stages (build, test, train, evaluate, package, sign, deploy) and release strategies (blue/green, canary, shadow, rollback with feature flags). 5) Detail runtime monitoring (data drift, prediction drift, performance, latency, error budgets, 24/7 alerts) and automated retraining triggers and approval workflows. 6) Cover secrets/PII handling, access control, and how you’d run an incident response and rollback within 15 minutes.
Quick Answer: This question evaluates MLOps and production machine learning engineering competencies, covering repository and environment management, model and data versioning, CI/CD/CD pipelines, runtime monitoring, security and governance, and reproducibility and incident response.
