MLOps¶

Reproducible inputs, tracked experiments, orchestrated training, safe rollout, and drift detection — the operational discipline that keeps a model trustworthy after it ships, not just accurate the day it was trained.

flowchart LR Store["Feature Store"] --> Track["Experiment Tracking<br/>& Model Registry"] Track --> Train["Model Training<br/>Pipelines"] Train --> Deploy["Model Deployment<br/>& Serving"] Deploy --> Monitor["Model Monitoring<br/>& Drift"] Monitor -.->|drift triggers retrain| Train

Topics¶

#	Topic	Status	Focus
01	Feature Store	Available	Reproducible model inputs across training and serving — offline/online stores, point-in-time correctness
02	Experiment Tracking & Model Registry	Coming soon	Versioning models/datasets/hyperparameters, promoting a registered model to production
03	Model Training Pipelines	Coming soon	Orchestrating training as a reproducible pipeline, retraining triggers, lineage
04	Model Deployment & Serving	Coming soon	Batch vs. real-time serving, canary/shadow rollout, rollback
05	Model Monitoring & Drift	Coming soon	Data drift, concept drift, performance decay, and retraining triggers

How to use this section¶

Feature Store is the foundation — it's what every other MLOps topic reads from or writes to. The remaining four topics follow the lifecycle of a model in production: track and register an experiment, train it through an orchestrated pipeline, deploy it safely, then monitor it for drift that feeds back into the next training run. This section is distinct from AI Evaluation, which covers LLM/agent-specific observability (hallucination, groundedness, prompt regressions) — MLOps here covers classic ML model lifecycle management.

Practice rule¶

Before deploying a model, confirm you can answer: what data trained it, what experiment produced it, and what signal will tell you it's drifting. If you can't answer all three, you have a model, not an operable one.

Part of Engineer Knowledge → AI Engineering.