Myths, Glossary, Cheat Sheet

Let's finish the technical part with three things worth saving. The most common myths about M L Ops, a glossary of the words you now know, and a cheat sheet of commands. Pause on any screen, and take a screenshot. These pages are made to be kept.

• Myths to forget
• Words to remember
• Commands to keep

Four myths about MLOps

First, the myths. Myth one, M L Ops needs a big platform. In truth, a small team can get most benefits, with Git, a pipeline, a tracker, and a container. Myth two, once the model is deployed, the work is done. In truth, that is where it starts, because models decay. Myth three, more complex tools mean better operations. Usually the opposite, simple, boring, tested systems win. Myth four, M L Ops is only for engineers. It also needs data scientists, who write testable code, and product people, who define success.

Glossary, part one

Now, the glossary. Drift, when production data moves away from training data. Skew, when training and serving data are prepared differently. Leakage, using information from the future. Pipeline, a chain of automated stages. Artifact, a file produced by a stage, such as data, or a model. Registry, the home of model versions. Champion, the model currently in use. And challenger, a candidate, that wants to replace it.

Glossary, part two

More words. Feature store, a shared catalogue of features, for training and serving. Canary, a release to a small share of traffic. Shadow mode, predicting silently, beside the live model. A B test, comparing two versions on real outcomes. S L O, a reliability target. Error budget, how much failure the target allows. And model card, a short document describing a model's purpose, data, limits, and owner. You can now follow most M L Ops conversations.

The MLOps cheat sheet

And the cheat sheet. Git versions code. D V C add, push, and repro version data, and run pipelines. M L flow tracks runs, and registers models. Pytest tests code, data, and models. Docker build and run package, and start the service. Curl calls the prediction endpoint. Kube control apply deploys to Kubernetes. And a drift script, run on a schedule, closes the loop. These eight tools, and a habit of automating, cover most of the work.

Summary