[reference] — from the roadmap.sh mlops roadmap. What MLOps is, and — the key framing — what it adds beyond the DevOps you already know.
What MLOps is
MLOps is the discipline of reliably building, deploying, and maintaining ML models in production. It’s DevOps applied to ML, and it borrows most of DevOps wholesale — CI/CD, containers, infrastructure-as-code, monitoring. So rather than re-explain those, this section cross-links the DevOps domain for the shared parts and focuses on what’s genuinely ML-specific.
What ML adds beyond ordinary DevOps
Traditional software has one thing that changes: code. ML systems have three things that change and interact — and that’s the entire source of MLOps’s extra complexity:
Traditional: code → build → test → deploy → monitor
ML: code + DATA + MODEL → ... → deploy → monitor for DRIFT → retrain
- Data is a versioned artifact. The same code trained on different data produces a different model. You must version data, not just code (DVC, feature stores). Git doesn’t handle multi-GB datasets.
- The model is a versioned artifact — a trained binary that must be tracked, registered, and rollback-able, tied to the exact data + code that produced it (model registry).
- “Correct” isn’t binary and it decays. Software is right or has a bug; a model is accurate enough until the world shifts and it silently degrades (drift). It needs retraining, not just patching.
- Testing is different. You can’t unit-test “is this prediction correct.” You test data quality, validate model performance against thresholds, and check for training-serving skew (serving).
The ML lifecycle
MLOps automates and monitors this loop, rather than treating training as a one-time event:
data collection → data validation → feature engineering → training
→ evaluation → model validation → deployment → monitoring
→ (drift detected) → retrain → redeploy ...
The goal is reproducibility (rebuild any model from versioned data+code+config), automation (retrain/deploy via pipelines, not by hand), and continuous monitoring (catch degradation before it hurts).
MLOps maturity
A useful way to place a team: from manual (notebooks, hand-deployed models, no monitoring — where most start) → automated training pipelines (retrain on a trigger) → full CI/CD for ML (automated retraining, validation, deployment, and monitoring closing the loop). You don’t need the top rung on day one; match the investment to how much the model matters and how fast its world changes.
What maps to DevOps (cross-links, not re-explained)
| MLOps needs | Lives in DevOps |
|---|---|
| Containerizing a model service | Docker / Kubernetes |
| Automated pipelines | CI/CD |
| Provisioning training/serving infra | Terraform/IaC |
| Metrics/logs/dashboards | Observability |
MLOps = these + data/model versioning + experiment tracking + drift monitoring + retraining. The next notes cover that ML-specific delta.
Related
- Experiment Tracking & Pipelines — versioning data/models/experiments
- Monitoring & Edge — drift and edge deployment
- DevOps — the shared foundation MLOps extends
Software engineering
Infrastructure & security
Computer science
Hardware & engineering