Artificial Intelligence & Data Science · 1.7

MLOps Engineer

Automate and operationalize ML workflows from training to deployment to monitoring.

6–8 months6 phases25 skills

Foundations

Linux administrationprocess management, networking (TCP/IP, DNS, firewalls), storage, systemd services, bash scripting for automation

Python

Pythonproduction scripting with logging, testing with pytest, packaging with pyproject.toml, virtual environments

Git

Gittrunk-based development, CI/CD concepts, branching strategies for ML experiments

Cloud fundamentals

Cloud fundamentalsAWS compute (EC2 with GPU instances — T4/A10G/A100), storage (S3), networking (VPC, security groups), IAM (roles with least-privilege for pipeline service accounts)

Phase 02 · 9h

Containerization & Orchestration

Docker advanced

Docker advancedmulti-stage builds to separate training and serving environments, GPU-enabled containers (NVIDIA CUDA base images with PyTorch CUDA support, verify with torch.cuda.is_available()), Docker Compose for local ML stacks (Jupyter + MLflow + PostgreSQL + MinIO), image optimization (pin versions, .dockerignore, cache pip installations)

Kubernetes for ML

Kubernetes for MLpods, deployments, services, ConfigMaps, Secrets, namespaces (ml-training, ml-serving, ml-monitoring), GPU scheduling with NVIDIA device plugin (resources: limits: nvidia.com/gpu: 1), Jobs for training runs, CronJobs for scheduled retraining, PriorityClasses (training lower priority than serving), Helm charts with configurable values (model version, replica count, GPU type)

Container registries

Container registriesECR (AWS), GCR (GCP), image tagging strategies for model versioning

Phase 03 · 13h

ML Pipeline Automation

Data versioning with DVC

Data versioning with DVCtrack large files alongside Git (dvc add creates .dvc metadata files, actual data stored in S3), define multi-stage ML pipelines in dvc.yaml (prepare → train → evaluate), run with dvc repro (executes only changed stages), experiments with dvc exp run --set-param (compare results with dvc exp show)

Pipeline orchestration with Kubeflow Pipelines (KFP)

Pipeline orchestration with Kubeflow Pipelines (KFP)define pipelines using KFP SDK v2 (@dsl.component for containerized steps, @dsl.pipeline for the workflow), caching (skip unchanged steps), lineage tracking, recurring runs; deploy on Vertex AI Pipelines (managed) or self-hosted on Kubernetes

Alternative

AlternativeMetaflow (by Netflix) — define pipelines as Python classes with @step decorators, @resources(gpu=1, memory=16000) for requirements, seamlessly scales from laptop to AWS Batch

Feature stores with Feast

Feature stores with Feastdefine Entity and FeatureView objects, materialize features from offline store (BigQuery) to online store (Redis), retrieve for training with get_historical_features (point-in-time correct), retrieve for serving with get_online_features, monitor feature freshness

Automated training

Automated trainingtrigger types (scheduled, performance-based when accuracy drops, data-driven when new labeled data is available), pipeline components (data validation → feature engineering → training with checkpointing → evaluation → model registration if improved → approval gate), hyperparameter optimization with Optuna integrated into the pipeline

Model registry

Model registryMLflow Model Registry (register models with versions, assign stages None→Staging→Production→Archived) and W&B Model Registry (link model version to training run, dataset version, and code commit); every model must have metadata (training data hash, Git commit, hyperparameters, evaluation metrics, known limitations)

Phase 04 · 15h

Model Serving & Monitoring

Model serving

Model servingTorchServe (model packaging with torch-model-archiver, worker management, dynamic batching), NVIDIA Triton (multi-framework support, GPU scheduling, model ensemble), vLLM for LLM serving (PagedAttention for efficient memory), batch prediction with Ray (@ray.remote for distributed inference), auto-scaling with Kubernetes HPA (scale based on requests_per_second or GPU utilization)

Deployment strategies

Deployment strategiesblue/green (instant traffic switch with rollback), canary (1% → 5% → 25% → 100% with monitoring at each stage), shadow deployment (new model runs alongside current, predictions logged but not served — validate before switching), A/B testing (route percentage of traffic to each model version, compare business metrics)

Monitoring — four pillars

Monitoring — four pillars(1) Infrastructure — CPU/GPU utilization, memory, disk with Prometheus + Grafana, (2) Application — latency p50/p95/p99, throughput, error rate, (3) Data — feature distribution drift vs training data with Evidently AI (Kolmogorov-Smirnov test for continuous features, Chi-squared for categorical, PSI with thresholds: <0.1 stable, 0.1-0.25 investigate, >0.25 retrain), (4) Model — prediction distribution shift, accuracy on recent data if ground truth available

Alerting

AlertingP1/Critical (serving down → PagerDuty, immediate), P2/Warning (drift detected → Slack, investigate in 24h), P3/Info (training completed → email digest)

SRE for ML

SRE for MLdefine SLIs (latency p99 < 200ms, availability > 99.9%, prediction quality > 85% accuracy), set SLOs, calculate error budgets, model rollback capability, disaster recovery (always keep previous model version ready)

Cost optimization

Cost optimizationspot/preemptible GPU instances for training (70-90% savings with checkpointing every N minutes), mixed precision training (fp16/bf16 halves memory, doubles throughput), right-sizing GPU instances, auto-scaling down to zero during off-peak

Phase 05 · 4h

Infrastructure & Security

Terraform modules for ML infrastructure

Terraform modules for ML infrastructuretraining cluster (EKS with GPU node groups, NVIDIA plugins, autoscaling), model registry (S3 bucket + DynamoDB metadata table + IAM roles), serving infrastructure (ECS/EKS deployment with ALB, auto-scaling, CloudWatch alarms)

Security

Securitymodel access control (IAM roles per environment), data encryption (at rest with KMS, in transit with TLS), audit logging (all model predictions, data access, configuration changes), compliance documentation for SOC 2 / HIPAA

Load testing with Locust

Load testing with Locustsimulate production traffic (from locust import HttpUser, task — define predict endpoint calls), measure system behavior under 10x, 50x, 100x normal load

Phase 06 · 18h

Capstone

Capstone Project 1

Capstone Project 1Full MLOps platform — Airflow data pipeline → Great Expectations validation → Feast feature store → Metaflow training pipeline with XGBoost + Optuna + MLflow → FastAPI serving on Kubernetes with HPA → Evidently drift monitoring → Prometheus/Grafana dashboards → GitHub Actions CI/CD → architecture documentation and runbooks

Capstone Project 2

Capstone Project 2Model serving platform with A/B testing — deploy two model versions simultaneously, route traffic with canary deployment, compare metrics, automated promotion of winner, auto-scaling on Kubernetes

Capstone Project 3

Capstone Project 3ML platform migration — given a legacy monolithic ML script, migrate to modern stack (DVC for data, MLflow for experiments, Docker for reproducibility, Kubernetes for serving, Evidently for monitoring), document before/after architecture, run both systems in parallel to validate, cut over

Ready to start MLOps Engineer?

Start this roadmap to track progress and generate courses for each skill. Login required only to begin.