MLOps: CI/CD & Retraining

The arc of MLOps: CI/CD & Retraining: a CI pipeline that goes green on a clean machine, CI that tests the data and not just the code, a gate that blocks a model that does not beat baseline, a retraining loop that can poison itself, and a shadow-then-canary rollout that ships a new version without betting the whole system on it.

Build a real CI pipeline that runs on every push, installs the locked deps, runs the tests, builds the artifact, and turns a clean machine's green check into evidence about your code.

Test data as a second input, not just code: assert schema, types, ranges, and nullability at the boundary, and catch the leakage that makes a model look brilliant in validation and fail in production — the failure no code-only test suite can see.

A green build does not mean a good model. Gate promotion on a metric that survives class imbalance rather than accuracy, and version the model as its artifact plus the data snapshot and seed that produced it — so it can be reproduced, audited, and rolled back.

Automatic retraining keeps a model up with drift and is also how it poisons itself: train on labels its own predictions influenced and each cycle gets worse while every metric looks fine. Build the loop with the guards that stop it degrading itself.

Ship a new model version without betting the whole system on it: shadow on the live stream, canary a bounded slice, then full, and standardize the path so months-long deploys stay three-day deploys.

A recap of MLOps: CI/CD & Retraining: a pipeline that goes green only on a clean machine, CI that tests data and gates the model on the right metric, versioned reproducible models, a retraining loop with guards against self-poisoning, and a shadow-then-canary rollout — and what you have built across the whole track.