✦ MLOps 90-Day Learning Path

What MLOps means

MLOps applies DevOps engineering discipline to the machine learning lifecycle — from data preparation and experiment tracking through model packaging, serving, monitoring, and retraining. It addresses the unique challenges of ML systems: data drift, model decay, reproducibility, and the dual ownership between data scientists and platform engineers. A mature MLOps practice ensures models in production are reliable, observable, and continuously improved.

Who should follow this path

• Data scientists who want to own their models in production
• ML engineers building model serving infrastructure
• Platform engineers supporting ML workloads on Kubernetes
• DevOps engineers adding ML pipeline expertise
• Software engineers at ML-driven product companies

Prerequisites

• Python proficiency and familiarity with scikit-learn or PyTorch
• Basic understanding of ML model training concepts
• Docker and containerization experience
• Familiarity with CI/CD pipelines
• Basic cloud platform experience (AWS, GCP, or Azure)

The 90-day plan

Daily study recommendation: 2-3 hours/day, six days a week. Consistency beats intensity — block the time in your calendar like a meeting.

Days 1–15: Foundation

• MLOps maturity levels and lifecycle stages
• ML pipeline components: data, features, training, evaluation, serving
• Reproducibility challenges in ML
• ML project structure and cookiecutter templates
• Version control for data, code, and models

Outcome: Structure an ML project for reproducibility with version-controlled data, code, and model artifacts.

Days 16–30: Core concepts

• Experiment tracking with MLflow
• Hyperparameter tuning with Optuna or Ray Tune
• Model registry: versioning and stage transitions
• Dataset versioning with DVC
• Comparing and selecting experiments systematically

Outcome: Track all experiments in MLflow, register the best model, and promote it through staging to production.

Days 31–45: Tools and workflows

• Feature stores with Feast or Tecton
• Feature engineering pipelines at scale
• Training pipelines with Kubeflow Pipelines or ZenML
• Containerizing ML training jobs
• Distributed training concepts with Horovod or Ray

Outcome: Build a feature store-backed training pipeline deployed on Kubernetes with Kubeflow.

Days 46–60: Hands-on projects

• Model serving with FastAPI, TorchServe, and BentoML
• Kubernetes-native serving with KServe (formerly KFServing)
• A/B testing and canary deployments for models
• Shadow mode deployment patterns
• Low-latency inference optimization (quantization, ONNX)

Outcome: Deploy a model as a production API with canary rollout and latency benchmarking.

Days 61–75: Advanced practices

• Data drift detection with Evidently AI
• Model performance monitoring and alerting
• Concept drift vs data drift detection methods
• Automated retraining triggers and pipelines
• Model explainability with SHAP and LIME

Outcome: Implement drift detection, automated alerting, and triggered retraining for a production model.

Days 76–90: Portfolio, interview & certification prep

• MLOps portfolio: end-to-end model pipeline project
• Preparing for Databricks ML Professional certification
• MLOps interview questions and system design
• ML platform metrics: training cost, inference latency, model accuracy SLAs
• Emerging topics: LLMOps, foundation model fine-tuning pipelines

Outcome: Present a complete MLOps pipeline project and be ready for ML engineer and MLOps engineer interviews.

Weekly outcomes at a glance

Phase	Outcome
Days 1–15	Structure an ML project for reproducibility with version-controlled data, code, and model artifacts.
Days 16–30	Track all experiments in MLflow, register the best model, and promote it through staging to production.
Days 31–45	Build a feature store-backed training pipeline deployed on Kubernetes with Kubeflow.
Days 46–60	Deploy a model as a production API with canary rollout and latency benchmarking.
Days 61–75	Implement drift detection, automated alerting, and triggered retraining for a production model.
Days 76–90	Present a complete MLOps pipeline project and be ready for ML engineer and MLOps engineer interviews.

Tools to learn

• MLflow
• Kubeflow Pipelines
• DVC
• Feast
• KServe
• BentoML
• Evidently AI
• ZenML
• Ray
• Optuna
• FastAPI
• Seldon Core

Labs to practice

• MLOps Pipeline Basics
• Kubernetes Deployment with Helm
• CI/CD with GitHub Actions
• Docker Build and Push

Mini projects

• Build an end-to-end MLflow + DVC + KServe pipeline from experiment tracking to production serving with drift monitoring
• Implement a Kubeflow Pipelines workflow for automated model retraining triggered by Evidently drift alerts
• Deploy a model A/B test on Kubernetes using KServe traffic splitting with latency and accuracy monitoring

Interview questions to prepare

1. What is the difference between a model registry and a feature store?
2. How do you detect and respond to data drift in a production model?
3. Explain the concept of shadow mode deployment for ML models.
4. How would you design a retraining pipeline that automatically triggers on performance degradation?
5. What is the difference between data drift and concept drift?
6. How do you ensure reproducibility of ML experiments across different environments?
7. Describe the architecture of a production-grade feature store.
8. What metrics would you track for an ML model in production?

Certification suggestions

• Databricks Certified Machine Learning Professional — Databricks
• AWS Certified Machine Learning Specialty — AWS
• Google Professional Machine Learning Engineer — Google Cloud
• Kubeflow Fundamentals — Linux Foundation

Browse the full certification registry for exam details and official links.

Free resources

• MLflow Documentation
• DVC Documentation
• Evidently AI Documentation
• ZenML Documentation
• Kubeflow Documentation

Related roadmaps

• MLOps Engineer Roadmap
• DataOps Engineer Roadmap
• AIOps Engineer Roadmap

Related tool categories

• MLOps Tools
• DataOps Tools
• Container Tools
• Kubernetes Tools
• Workflow Orchestration Tools