MLOps Pipeline

the MLOps pipeline industrializes the model lifecycle “data → training → evaluation → deployment” with four pillars: ① CI/CD for ML — code, data and models are all artifacts with their own build/test/release pipelines, so a data change or code merge auto-triggers training smoke tests; ② experiment tracking — MLflow / W&B record hyperparameters, metrics, code commit, dataset version and artifact links per run so every experiment is comparable in one coordinate system; ③ reproducibility — fix random seeds, lock data versions and environments, and write a fingerprint of the run into the record $h = H(\text{data} \parallel \text{config} \parallel \text{seed})$ so anyone can reproduce it with one command; ④ model registry & canary release — the registry manages candidate stages (dev/candidate/prod), versions and metadata, while traffic share $p$ ramps from 1% to 100% and rolls back on anomaly.

moving training from “one person's notebook” to a multi-person production workflow; interviews ask how ML CI/CD differs from software CI/CD, how tracking keeps comparisons fair, how to debug irreproducibility, and how canary release relates to A/B testing.

without a pipeline, experiments are irreproducible (different machine, different numbers), candidates cannot be compared fairly, and shipping a checkpoint by hand causes incidents; CI/CD surfaces problems before merge via data validation and training smoke tests; the registry plus canary release cuts rollback from hours (retrain + redeploy) to minutes and shrinks blast radius from 100% of traffic to a 1% observable window that can be reverted instantly.

How does CI/CD for ML differ from traditional software CI/CD? Why must data and models be managed as artifacts?

Which fields must experiment tracking (MLflow/W&B) record at minimum, and how does it keep cross-experiment comparison fair?

How do you make training reproducible? Is fixing the seed enough (data version/env/framework/hardware)?

Model registry vs model repository? How to design canary traffic ramps and rollback?

Given a notebook training run, how does a colleague reproduce identical results on another machine? Where do you look when it fails?