THE CORE AIM OF VOCABULARY MASTERY · MLOPS VOCABULARY · DATA → TRAIN → REGISTER → DEPLOY → MONITOR
MLOps vocabulary is the language used to describe the operational lifecycle of machine-learning models from experiments and training through deployment, monitoring and retraining. Terms such as experiment tracking, model registry, feature store, training pipeline, inference, drift, model monitoring and retraining matter because production machine learning must manage both software and changing data.
The core aim of vocabulary mastery for MLOps vocabulary is model-lifecycle clarity. Learners should be able to explain which data and code produced a model, how that model was evaluated and versioned, how it reached production, what signals are monitored and what triggers retraining or rollback.
This page is the MLOps Vocabulary owner inside the eduKateSG Vocabulary hub. For model concepts, use Machine Learning Vocabulary. For software delivery, use DevOps Vocabulary.
Central proposition: MLOps vocabulary is mastered when a production prediction can be traced back through model version, training run, data, code and evaluation evidence.
The MLOps Vocabulary Router
- Experiment: run, parameter, metric, artifact.
- Data: dataset version, feature, feature store, lineage.
- Model: checkpoint, registry, version, approval.
- Deploy: endpoint, batch inference, online inference, rollout.
- Monitor: latency, drift, quality, data distribution.
- Maintain: retrain, rollback, champion, challenger.
MLOps and DevOps Are Related but Different
DevOps focuses on reliable software delivery and operations. MLOps adds concerns unique to learned models: training data, experiments, model versions, feature consistency, drift and retraining.
A Worked Example: Model Registry
A model registry stores or tracks model versions together with metadata such as training run, evaluation results and deployment status. It helps teams know exactly which model is approved or serving.
A Worked Example: Drift
Data drift means the distribution of incoming data changes relative to a reference period. Concept drift means the relationship between inputs and desired outcomes changes. Either can reduce model usefulness even when software remains healthy.
Online and Batch Inference
Online inference produces predictions in response to requests, usually with tighter latency requirements. Batch inference scores groups of records on a schedule or job. Operational vocabulary differs because throughput, freshness and failure handling differ.
Retraining and Rollback
Retraining produces a new model using updated data or procedures. A rollback returns serving to a previously approved version when a new model performs poorly or causes operational problems.
How to Learn MLOps Vocabulary
- Track one experiment end to end.
- Version data, code and model artifacts.
- Register the chosen model.
- Deploy it to a safe endpoint.
- Monitor latency and prediction quality.
- Simulate drift conceptually.
- Practise retraining and rollback decisions.
Common Mistakes
Treating model deployment as the end
Repair: include monitoring, drift and retraining.
Versioning code but not data
Repair: preserve training-data lineage.
Monitoring uptime but not model quality
Repair: include predictive and data-distribution signals.
Frequently Asked Questions
What is MLOps vocabulary?
It is the language used for machine-learning experiments, model registries, deployment, monitoring, drift and retraining.
What is a model registry?
It is a system for tracking model versions, metadata, evaluation and lifecycle status.
What is data drift?
It is a change in the statistical distribution of incoming data relative to a reference distribution.
The MLOps Vocabulary Standard
Mastery means tracing a production model through its data, training run, evaluation, registry, deployment and monitoring history.
That is the standard: operational ML language that makes learned systems reproducible and maintainable.