MLOps is the engineering that gets a machine-learning model out of a notebook and keeps it working in production: reproducible training, versioned data and models, deployment, monitoring and retraining. A model is code plus data plus configuration, and it decays as the world changes, so the pipeline around it matters more than any single run.
Key concepts
- Reproducibility: fixed seeds, pinned dependencies, recorded configuration and versioned data, so a run can be repeated.
- Pipelines: data validation, feature computation, training, evaluation and packaging as automated steps.
- Deployment patterns: batch scoring, online serving, shadow and canary releases, with a fast rollback.
- Monitoring: latency and errors, but also input drift, prediction drift and quality against delayed labels.
- Training-serving skew: features computed differently offline and online give silent failures.
Vocabulary
- Feature store: a shared system that serves the same feature definitions for training and serving.
- Model registry: a catalogue of model versions with their metrics and status.
- Data drift: the input distribution in production moving away from the training one.
- Concept drift: the relation between inputs and the right answer changing over time.
What to study, in order
- Git, containers and a CI pipeline, which everything else rests on.
- Take a small model, script its training end to end, and make two runs give identical results.
- Version the data and log every experiment with parameters and metrics.
- Serve the model behind an HTTP API with input validation, tests and a latency measurement.
- Add monitoring for drift, simulate a shift in the inputs, and write the retraining and rollback procedure.
Related tags
#machine-learning #devops #ci-cd #observability #data-engineering