CD for ML: Promote, Don't Re-train

A different shipping unit

Classic CD ships the binary of your code. ML CD ships the trained model artefact (or a registry version pointer). The two are connected:

• merging code triggers a training pipeline (Argo, Vertex Pipelines, GitHub Actions, Airflow);
• the pipeline produces a candidate model version;
• a metric gate decides whether to register and promote;
• a canary deployment slowly shifts traffic.

Crucially, the same trained artefact moves through all environments. You don't 'rebuild' the model for prod — rebuilding for each environment is exactly how you introduce skew and lose reproducibility.

Manufacturing analogy: the batch number travels through every checkpoint (QC, packaging, shipping). Nobody re-makes the batch downstream — they promote the same batch through gates. ML CD works the same way.