Catalog
technology#Machine Learning#Platform#Cloud#MLOps / DevOps

Kubeflow

Kubeflow is an open-source platform for orchestrating and deploying machine learning workflows on Kubernetes.

Kubeflow is an open-source platform for developing, orchestrating, and deploying machine learning workloads on Kubernetes.
Established
High

Classification

  • High
  • Technical
  • Technical
  • Advanced

Technical context

Kubernetes (CRDs, operators)Cloud storage (S3, GCS), object storageMonitoring stacks (Prometheus, Grafana)

Principles & goals

Kubernetes-native integrationRepeatability and reproducibilitySeparation of training and serving
Build
Enterprise, Domain, Team

Use cases & scenarios

Compromises

  • Misconfigurations can lead to resource waste
  • Security and access controls must be carefully managed
  • Dependency on Kubernetes ecosystem and provider APIs
  • Version models and pipelines
  • Define resource limits and quotas
  • Automate tests and validation within pipelines

I/O & resources

  • Kubernetes cluster with sufficient resources
  • Containerized training images
  • Access to distributed storage (S3, GCS, NFS)
  • Trained model artifacts and versions
  • Pipeline metadata and logs
  • Production endpoints and monitoring metrics

Description

Kubeflow is an open-source platform for developing, orchestrating, and deploying machine learning workloads on Kubernetes. It provides reusable pipelines, model serving, distributed training support, monitoring, and integrations with cloud Kubernetes offerings. Kubeflow enables scalable, cloud-ready MLOps workflows but requires Kubernetes and platform engineering expertise.

✔

  • Standardized pipelines for reproducibility
  • Scalability via Kubernetes orchestration
  • Wide integration with cloud Kubernetes providers

✖

  • Requires solid Kubernetes expertise
  • Operational and upgrade complexity
  • Not all ML tools integrated out-of-the-box

  • Pipeline run time

    Duration from start to completion of a training pipeline to measure efficiency.

  • Cost per training run

    Monetary expenses for resource consumption during a training run.

  • Inference latency and throughput

    Measures to evaluate serving performance in production.

Large e-commerce company

Uses Kubeflow to automate recommendation models and for A/B testing in production.

Image processing research institute

Uses Kubeflow for distributed training of large neural networks on a GPU cluster.

Financial services provider

Uses Kubeflow for reproducible model audits, governance and feature traceability.

1

Cluster preparation: configure resources, RBAC, storage

2

Install Kubeflow components and integrations

3

Set up example pipelines, tests and monitoring

⚠️ Technical debt & bottlenecks

  • Custom operators using deprecated APIs
  • Insufficiently documented pipeline variants
  • Hardcoded configurations instead of configurable manifests
GPU/TPU resource availabilityNetwork I/O and storage I/OPlatform upgrade complexity
  • Running training jobs without quotas leads to cluster disruption
  • Storing sensitive data in unencrypted storage
  • Using outdated pipelines without tests
  • Underestimating operational costs for large GPU clusters
  • Ignoring RBAC and network policies
  • Compatibility issues during component upgrades
Kubernetes administrationContainerization and CI/CDMachine learning workflow design
Scalability for training and inference loadsRepeatable CI/CD-like MLOps processesIntegration with Kubernetes ecosystem and cloud providers
  • • Requires Kubernetes cluster and appropriate RBAC configuration
  • • Dependency on third‑party operators for specific functionality
  • • Compliance requirements may limit cloud deployment