MLOps Discipline

1 module is currently being reworked. Watch this section over the next few days.

Discipline Track | 12 Modules | ~10 hours total

Overview

MLOps brings engineering rigor to machine learning. Most ML projects fail not because of bad models, but because teams can’t operationalize them. Data scientists build prototypes; MLOps turns them into production systems.

This track covers the complete ML lifecycle—from experiment tracking and feature stores to data versioning, model serving, monitoring, and automated pipelines—giving you the skills to deploy and maintain ML systems at scale.

Prerequisites

Before starting this track:

Observability Theory Track — Monitoring fundamentals
Basic machine learning concepts (training, inference, models)
Python programming experience
Understanding of CI/CD concepts
Kubernetes basics (helpful but not required)

Modules

#	Module	Complexity	Time
5.1	MLOps Fundamentals	`[MEDIUM]`	35-40 min
5.2	Feature Engineering & Stores	`[COMPLEX]`	40-45 min
5.3	Model Training & Experimentation	`[COMPLEX]`	40-45 min
5.4	Model Serving & Inference	`[COMPLEX]`	40-45 min
5.5	Model Monitoring & Observability	`[COMPLEX]`	40-45 min
5.6	ML Pipelines & Automation	`[COMPLEX]`	40-45 min
5.7	Data Versioning with DVC	`[COMPLEX]`	50-60 min
5.8	Great Expectations Data Quality	`[COMPLEX]`	50-60 min
5.9	ML Repository Hygiene	`[COMPLEX]`	45-55 min
5.10	Production Model-Serving Traffic Patterns	`[COMPLEX]`	55-65 min
5.11	Drift-Triggered Auto-Retraining Loop	`[COMPLEX]`	60-70 min
5.12	CML for ML CI	`[COMPLEX]`	55-60 min

Learning Outcomes

After completing this track, you will be able to:

Understand MLOps maturity — From notebooks to automated pipelines
Build feature stores — Ensure consistency between training and serving
Track experiments — Reproduce results, compare approaches systematically
Deploy models — KServe, canary deployments, A/B testing
Monitor ML systems — Detect drift, track performance without labels
Automate pipelines — Kubeflow, continuous training, CI/CD for ML
Version data and models — Use DVC to connect Git commits, data hashes, model artifacts, and metrics
Gate data quality — Use Great Expectations to validate schema, completeness, and distribution contracts before training or serving
Maintain clean ML repositories — Keep data, models, notebooks, dependencies, hooks, and CI policies reviewable without bloating Git history
Control serving exposure — Use canary, A/B, shadow, mirroring, and bandit patterns to promote models with measurable rollback and cost controls
Close the model lifecycle loop — Close the model lifecycle loop with drift-triggered automated retraining, gated promotion, and forensic rollback
Surface model review evidence — Surface model metric deltas, validation reports, and deployment health back into the Git PR review surface using CML on GitHub Actions or GitLab CI.

Key Concepts

The ML Lifecycle

┌─────────────────────────────────────────────────────────────────┐
│                      ML LIFECYCLE                                │
│                                                                  │
│  DATA                 EXPERIMENTATION           PRODUCTION       │
│  ┌──────────┐        ┌──────────┐            ┌──────────┐       │
│  │  Data    │        │  Model   │            │  Model   │       │
│  │ Ingestion│───────▶│ Training │────────────▶│ Serving  │       │
│  └────┬─────┘        └────┬─────┘            └────┬─────┘       │
│       │                   │                       │              │
│  ┌────▼─────┐        ┌────▼─────┐            ┌────▼─────┐       │
│  │  Data    │        │  Model   │            │  Model   │       │
│  │Validation│        │Validation│            │Monitoring│       │
│  └────┬─────┘        └────┬─────┘            └────┬─────┘       │
│       │                   │                       │              │
│  ┌────▼─────┐        ┌────▼─────┐            ┌────▼─────┐       │
│  │ Feature  │        │  Model   │            │ Trigger  │       │
│  │  Store   │        │ Registry │            │ Retrain  │◀──────┘
│  └──────────┘        └──────────┘            └──────────┘       │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Core Principles

Reproducibility — Every training run must be reproducible
Automation — Automate everything from training to deployment
Versioning — Version code, data, AND models
Monitoring — ML systems fail silently; monitor everything
Continuous Training — Models degrade; keep them fresh

MLOps vs DevOps

Aspect	DevOps	MLOps
Artifact	Code	Code + Data + Model
Testing	Unit, integration	+ Model validation, drift tests
Versioning	Git	Git + DVC/MLflow
Monitoring	Infrastructure	+ Data quality, model performance
CI/CD	Build, test, deploy	+ Train, validate, serve

Tools Covered

Category	Tools
Experiment Tracking	MLflow, Weights & Biases, Neptune
Feature Stores	Feast, Tecton, Hopsworks
Model Serving	KServe, Seldon Core, BentoML, TorchServe
Pipeline Orchestration	Kubeflow Pipelines, Apache Airflow, Argo
Monitoring	Evidently, WhyLabs, Arize, NannyML
Hyperparameter Tuning	Optuna, Katib, Ray Tune
Platforms	Kubeflow, SageMaker, Vertex AI, Databricks

Study Path

Module 5.1: MLOps Fundamentals
     │
     │  Why ML is different, maturity levels
     ▼
Module 5.2: Feature Engineering & Stores
     │
     │  Training/serving skew, Feast
     ▼
Module 5.3: Model Training & Experimentation
     │
     │  MLflow, HPO, reproducibility
     ▼
Module 5.4: Model Serving & Inference
     │
     │  KServe, deployment patterns
     ▼
Module 5.5: Model Monitoring & Observability
     │
     │  Drift detection, Evidently
     ▼
Module 5.6: ML Pipelines & Automation
     │
     │  Kubeflow, CI/CD for ML
     ▼
Module 5.7: Data Versioning with DVC
     │
     │  Git + DVC metadata, remotes, reproducible data pipelines
     ▼
Module 5.8: Great Expectations Data Quality
     │
     │  Data contracts, checkpoints, Data Docs, K8s validation Jobs
     ▼
Module 5.9: ML Repository Hygiene
     │
     │  src layout, ignore policy, lock files, notebook discipline, pre-commit gates
     ▼
Module 5.10: Production Model-Serving Traffic Patterns
     │
     │  KServe canary, Istio A/B, shadow, mirroring, bandits, cost controls
     ▼
Module 5.11: Drift-Triggered Auto-Retraining Loop
     │
     │  Drift signals, Argo triggers, retraining DAGs, validation gates, rollback
     ▼
Module 5.12: CML for ML CI
     │
     │  PR comments, DVC metric deltas, validation reports, runner cost controls
     ▼
[Track Complete] → ML Platforms Toolkit

Before: Observability Theory — Monitoring foundations
Related: IaC Discipline — Infrastructure provisioning for ML platforms
Related: DevSecOps — Security for ML pipelines
After: ML Platforms Toolkit — Hands-on implementations
After: IaC Tools Toolkit — Terraform modules for ML infrastructure

“A model is only as good as the system that serves it. MLOps is that system.”