Glossary¶

This glossary defines key terms as they are used within this project. Definitions are pragmatic and system-specific.

MLOps¶

A set of practices for building, deploying, monitoring, and maintaining machine learning systems in production.

A tool used to version datasets and pipeline stages, ensuring reproducibility of experiments.

A platform for tracking experiments, storing artifacts, and managing model versions via a registry.

A centralized service that stores trained models, their metadata, and their lifecycle stages (e.g. staging, production).

A principle stating that the same feature logic and model artifacts must be used during training and online inference.

A formal definition of expected data schema and quality constraints, validated using tools such as Great Expectations.

The use of information during training that would not be available at prediction time, leading to overly optimistic evaluation results.

A minimal, deterministic sequence of steps that reproduces the full ML lifecycle from data to inference.

A low-latency inference mode where predictions are returned immediately within the request lifecycle.

A background inference mode using message queues and workers, suitable for heavier or batch workloads.

A change in data or model behavior over time that can degrade model performance.

A target level for service reliability or performance, such as latency or error rate.

Operational documentation describing how to diagnose and resolve common failures in a production system.