Why MLOps matters at Cloudflare
Cloudflare has relied on machine learning and AI for core services since its early days, and that experience has produced a body of practice around running AI deployments at scale. Models drive the WAF attack score that screens malicious payloads, power bot management, support data science work in customer support, and identify threats across the global network. Workers AI now abstracts much of the inference side away for developers; the focus here is the other half of the equation — training.
With ML models occupying a crucial position — experimentation, training and deployment alike — the MLOps team worked with Cloudflare's data scientists to implement a set of best practices.
Notebook environments
Most data scientists begin with an environment for data exploration, feature engineering and model experiments, and Jupyter Notebooks are the standard tool for the job. To make notebooks scalable and collaborative, Cloudflare deploys JupyterHub on Kubernetes, which lets the team manage resources for groups of data scientists and give each a suitable development environment. Teams tailor their environments with pre-installed libraries and per-project configuration.
The notebook space keeps evolving with open source work, including:
nbdev— a Python package that improves the notebook experienceKubeflow— the Kubernetes-native CNCF project for machine learningdeployKF— ML platforms on any Kubernetes cluster, with centralized configs, in-place upgrades, and support for tools such as Kubeflow, Airflow and MLflow
GitOps and delivery
Cloudflare's MLOps platform adopts GitOps for continuous delivery and infrastructure management. Git is the single source of truth for declarative infrastructure and application configuration, which makes deployment and management automated, auditable and consistent with the collaboration that ML workflows demand.
A data scientist defines the desired state of infrastructure and applications. With ArgoCD and model templates, that normally custom work reduces to a pull request to add new applications; Helm charts and Kustomize are both supported for configuring different environments and jobs. ArgoCD then runs the continuous delivery process, continuously checking that the live state stays in sync with the Git repository.

The roadmap calls for migrating the platform — JupyterHub included — to Kubeflow, which simplifies development, deployment and management of notebooks and end-to-end pipelines on Kubernetes. The team considers deployKF, which provides distributed configuration management across multiple Kubeflow components and beyond, the best way to run it.
Curated project templates
Starting with the right structure separates a project that succeeds from one that stalls. Cloudflare maintains a set of production-ready data science repositories, each built around an example model and deployed through production to keep them a stable foundation. One Makefile command creates a new CI/CD project in the Git repository of the user's choosing. The template utility packages match those used in the Jupyter Notebooks, with connections to R2 / S3 / GCS buckets and D1 / Postgres / Bigquery / Clickhouse databases. The templates are not yet public; the team plans to open source them.
- Training template — a base for any model, configured for ETL from any data source, with helpers for feature engineering and experiment tracking with model metadata, plus a choice of DAG orchestration — Airflow or Argo Workflows — to productionalize the pipeline.
- Batch inference template — schedules models for consistent batch and micro-batch results, configurable for Airflow or Argo Workflows.
- Stream inference template — real-time inference as a Kubernetes microservice built on FastAPI, so teams run inference with familiar Python in a container. It ships REST interactive documentation via Swagger and integrates Cloudflare Access authentication tokens in terraform.
- Explainability template — dashboards that illuminate model type and experiments, surfacing values such as time window F1 score and the drift of features and data over time, using tools like Streamlit and Bokeh.
Orchestration choices
Directed Acyclic Graphs (DAGs) weave the steps from data to model and model to inference into a consistent pipeline across many data versions and model versions. Cloudflare lets each team's preference decide among several approaches:
- Apache Airflow — the standard DAG-based composer, with a large community, extensive plugin support, integration with many systems and a web UI for monitoring complex task sequences. It runs any data or ML workflow.
- Argo Workflows — Kubernetes-native, with intuitive YAML workflow definitions and strength in microservices-based workflows. Container support delivers scalability and reliability for organizations rooted in Kubernetes; it also runs any data or ML workflow.
- Kubeflow Pipelines — tailored to ML workflows, addressing data preparation, model training and deployment, with emphasis on collaboration, reusability and versioning inside the Kubeflow ecosystem.
- Temporal — oriented toward long-running, stateful workflows and resilient, event-driven applications, preserving workflow state for efficient recovery from failures.
The choice comes down to the team and the use case. All are open source, so the only real constraint is support for different styles of work — an investment the team considers worthwhile.
Hardware and metrics
Optimal performance requires understanding workloads to supply effective hardware, balancing enablement of data scientists against utilization. Every workload differs, so each use case is tuned for the capabilities of GPUs and CPUs — GPUs for core datacenter workloads and edge inference deliver speed and efficiency, while observability and metrics consumed by Prometheus let the team track orchestration, optimize performance, maximize utilization and stay within a Kubernetes-native experience.
Driving adoption
Adoption is one of the hardest steps in any MLOps journey. At Cloudflare the process started years ago, when teams each built their own ML solutions separately; as those grew, work risked becoming isolated, and other teams with ML potential lacked data science expertise. MLOps stepped in to streamline and standardize processes across teams and to bring potential new projects to data science teams for ideation and discovery.
The greatest success comes from helping projects get started and shaping pipelines for it. Shared components — notebooks, orchestration, data versioning with DVC, feature engineering with Feast, and model versioning with MLflow — enable teams to collaborate directly.
What comes next
Data science keeps evolving Cloudflare's business and its customers' businesses. The company improves its own products with models and has built AI infrastructure to secure applications, including those built with AI, and to deliver AI over its network. Partnerships with machine learning companies aim to make it easier for the data science community to deliver real value from data. The invitation is to join the Cloudflare community in bringing modern software practices and tools to data science.



