5 collocation exercises on deploying and operating ML models.
0 / 5 completed
1 / 5
After training, you ___ a model in the registry.
You register a model — recording a versioned artifact in a model registry (like MLflow) with its metadata and lineage. Register collocates with model and version. Enrol off, log up and file out are not idiomatic. Registering models gives you a single source of truth for which version is staged or in production, enabling reproducibility and controlled promotion between environments.
2 / 5
In production, the system will ___ predictions.
The system serves predictions — exposing the model behind an API or batch job so applications can request inferences. Serve is the standard term (model serving, serve inference). Dish off, deliver up and supply out are not the collocation. Serving concerns latency, scaling and reliability of inference, which are distinct engineering challenges from training the model in the first place.
3 / 5
To test a new model on live traffic safely, you ___ it.
You shadow-deploy a model — running it alongside production on real traffic but without serving its results to users, so you can compare safely. Shadow deployment (or shadowing) is the established term. Ghost-run off, mirror up and echo out are invented. Shadow deploying lets you validate a new model's behaviour and performance under genuine load before committing to it.
4 / 5
Over time, you must ___ for drift.
You monitor drift — tracking whether input distributions or model accuracy degrade as the world changes (data drift, concept drift). Monitor collocates with drift, performance and metrics. Watch off, scan up and eye out are not idiomatic. Monitoring drift catches silent model decay; without it, a model that worked at launch can quietly become inaccurate as real-world patterns shift.
5 / 5
If a new model underperforms, you ___.
You roll back — reverting to the previous known-good model version when a deployment causes problems. The phrasal verb roll back collocates with deploy, model and release. Reverse off, undo up and revert out are not idiomatic. A fast, reliable rollback path is essential in ML ops, because monitoring may reveal a regression only after the new model is live.