Learn CI/CD for ML

CI/CD for ML automates the path from change to safe release. It includes normal software checks, but also data checks, training pipeline checks, model evaluation gates, artifact registration, canary deployment, and rollback. The goal is to make model shipping repeatable without making it reckless.

Extra gates for ML

• Code gates: unit tests, integration tests, type checks, linting, security scans, and container build.
• Data gates: schema validation, missingness limits, freshness, leakage checks, duplicate checks, and bias checks.
• Training gates: pipeline reproducibility, resource limits, deterministic seeds where possible, and artifact creation.
• Evaluation gates: global metrics, slice metrics, calibration, robustness, latency, memory, and cost.
• Governance gates: model card, approval, risk tier, audit fields, and production owner.
• Release gates: canary health, SLOs, drift watch, business guardrails, and rollback automation.

promote model if:
  tests pass
  data checks pass
  metric_delta >= threshold
  critical slices do not regress
  latency/cost stay within budget
  rollout health stays green

Promotion policy should be explicit, not vibes.