21-Step MLOps Roadmap: the Comments Get Stuck at Step 13

Someone on X posted an MLOps roadmap for developers. The author is Dhanian, and the link to the original post is at the end. It uses ASCII to string 21 stages into a single line, starting from machine learning fundamentals, Python, and software engineering, all the way through LLMOps and end-to-end pipelines. For a general reader, MLOps can be understood as the whole set of engineering actions that take a model from a notebook to production. Model training is only the middle part; before and after it is all data, versioning, deployment, monitoring, and governance.

The main structure of the roadmap is roughly as follows:

1. Machine Learning Fundamentals: supervised learning, unsupervised learning, deep learning.
2. Python for MLOps: Python, NumPy, Pandas, Scikit-learn.
3. Software Engineering: Git, GitHub, testing, clean code, APIs.
4. Data Engineering: data collection, cleaning, validation, transformation.
5. Data Versioning: DVC, Git LFS, dataset management.
6. Experiment Tracking: MLflow, Weights & Biases, Comet ML.
7. Model Development: training, validation, hyperparameter optimization.
8. Model Versioning: MLflow Registry, model artifacts, metadata.
9. Model Packaging: Python packages, Docker, model containers.
10. Model Serving: FastAPI, Flask, REST API, gRPC.
11. Workflow Orchestration: Airflow, Prefect, Kubeflow, Dagster.
12. Machine Learning CI/CD: GitHub Actions, GitLab CI, Jenkins, automated testing.
13. Infrastructure and Cloud: AWS, Azure, GCP, Kubernetes, Terraform.
14. Model Monitoring: Prometheus, Grafana, Evidently AI, WhyLabs.
15. Model Drift Detection: data drift, concept drift, performance monitoring.
16. Feature Store: Feast, feature engineering, feature serving.
17. Model Optimization: quantization, pruning, ONNX, TensorRT, batching.
18. Model Governance: model lineage, reproducibility, interpretability, auditing.
19. LLMOps and Generative AI: LLM deployment, RAG, vector databases, evaluation, guardrails.
20. Production MLOps Projects: prediction APIs, recommendations, fraud detection, image classification, NLP pipelines, real-time inference.
21. End-to-End MLOps Pipeline: data, training, evaluation, versioning, deployment, monitoring, ultimately leading to production MLOps engineer.

The comment section is more direct than the roadmap. One commenter wrote in a reply that once you get to around step 13, you stop doing much ML and become the person who owns a Kubernetes cluster. That complaint is very accurate. Step 13 is infrastructure and cloud, with Kubernetes and Terraform sitting right there; the work starts shifting from tuning models to managing clusters, configuring cloud resources, and maintaining deployment pipelines. The roadmap is not exaggerating—it is just writing out the reality of the role: the further you go, the lower the share of model code and the higher the share of platform and operations work. Job postings say ML, but on day one you look at YAML; this kind of mismatch is not new in MLOps roles.

For developers, this checklist has several practical points. Data versioning at step 5, experiment tracking at step 6, model versioning at step 8, monitoring at step 14, and drift detection at step 15 are all unavoidable once a model goes to production. Without them, training results are not reproducible, and when online performance drops, it is hard to pinpoint why. Step 19 singles out LLMOps—RAG, vector databases, evaluation, and guardrails—showing that generative AI has pushed the boundaries of MLOps further out. Steps 20 and 21 are good for verification: can you build prediction APIs, recommendations, and fraud detection, and can you connect data, training, evaluation, versioning, deployment, and monitoring into a repeatable pipeline?

This roadmap can be used as a checklist, not necessarily as a learning order. The real dividing line is step 13. If you only want to train models, you can call it done by step 8; if you want to get models into production, after step 13 is where the bills and on-call start. The original post is from Dhanian: link. Kunal’s reply is also under the original post.