● Beta — KubeEra is in active development. Profiles are live; AI-perception measurement is rolling out.
KubeEra
K

Kubeflow Pipelines

AI Infrastructure
Open source · Apache-2.0

Kubeflow Pipelines (KFP) is a platform for building, deploying, and managing multi-step ML workflows as containerized pipelines on Kubernetes.

updated 2026-07-01

What it is

Kubeflow Pipelines (KFP) is a platform for building, deploying, and managing multi-step ML workflows as containerized pipelines on Kubernetes. It gives you a Python SDK to define DAGs of components (each a container), a pipeline engine to orchestrate them (Argo Workflows under the hood, historically), a UI for tracking runs and artifacts, and metadata tracking for lineage. It's one of the core subprojects under the broader Kubeflow umbrella, focused specifically on the "orchestrate my training/eval/deploy steps" problem rather than serving or notebooks.

Who builds it and why

Google originally spun this out of internal ML infra work when Kubeflow launched, and it's remained heavily Google-influenced in design (Vertex AI Pipelines is built on the same KFP SDK/IR). Today it's a community project under the Kubeflow org with contributions from Google, Canonical, IBM, and various platform teams who run ML on Kubernetes and needed pipeline orchestration that wasn't Airflow bolted onto a cluster. 556 contributors is a real number for a project this scoped — this isn't a single-vendor pet project, but Google's fingerprints (and roadmap influence) are still visible.

Production readiness signal

Active: last commit dated 2026-07-01, current release 2.16.1, which tells you this is a maturing, iterating codebase, not abandonware. 4,164 GitHub stars is modest for a project of this age and adoption level — lower than you'd expect given how embedded KFP is in the Kubeflow ecosystem, which suggests visibility/marketing lags actual usage. CNCF maturity level is not publicly available (Kubeflow as a whole is a CNCF incubating project, but KFP-specific status isn't separately published). Apache-2.0, Python — standard, no licensing surprises. Treat "production ready" here as "yes if you're already running Kubernetes and comfortable with Kubeflow's operational overhead," not "yes out of the box for a five-person team."

Who should use this

  • Teams already running Kubeflow (or planning to) who need pipeline orchestration as part of that stack.
  • ML platform teams building internal MLOps tooling on Kubernetes who want a Python-native SDK for defining pipelines with strong artifact/metadata tracking baked in.
  • Orgs with existing Argo Workflows expertise — KFP's execution layer will feel familiar to debug.
  • Teams that need pipeline portability between on-prem Kubernetes and Vertex AI Pipelines (same SDK, same pipeline IR).

Who should NOT use this

  • Small teams without dedicated platform/infra engineers. KFP requires standing up and maintaining Kubeflow (or at minimum KFP standalone) on Kubernetes — that's real operational burden, not a pip install.
  • Anyone whose orchestration needs are general-purpose data pipelines, not ML-specific — use Airflow, Dagster, or Prefect instead; KFP's abstractions (components, artifacts, ML metadata) add friction for non-ML DAGs.
  • Teams not already on Kubernetes. Standing up a cluster just to run KFP is backwards; if you're cloud-native on GCP already, Vertex AI Pipelines gets you the same SDK without cluster ops.
  • Anyone needing low-latency or real-time orchestration — KFP is batch-oriented, built for training/eval workflows, not streaming or online inference chains.
  • Teams wanting a stable, rarely-changing API surface. Kubeflow's ecosystem (KFP included) has had real backward-compatibility churn across major versions (v1 to v2 SDK was a meaningful rewrite).

Alternatives

  • Airflow — the incumbent general-purpose orchestrator; better ecosystem/integrations, worse for ML-specific artifact/metadata tracking.
  • Argo Workflows — the engine KFP is built on; use directly if you don't need KFP's ML-specific abstractions and want less overhead.
  • Vertex AI Pipelines — same SDK as KFP, fully managed on GCP; use this if you don't want to operate Kubernetes yourself.

Pricing

Fully open source, Apache-2.0. No paid tier, no vendor lock-in from the project itself. Your cost is the Kubernetes cluster and engineering time to run and maintain it — that's the real "price" of KFP, and it's not small.

Frequently asked

What is Kubeflow Pipelines?+
Kubeflow Pipelines (KFP) is a platform for building, deploying, and managing multi-step ML workflows as containerized pipelines on Kubernetes.
Who builds Kubeflow Pipelines?+
Google originally spun this out of internal ML infra work when Kubeflow launched, and it's remained heavily Google-influenced in design (Vertex AI Pipelines is built on the same KFP SDK/IR).
Is Kubeflow Pipelines production ready?+
Active: last commit dated 2026-07-01, current release 2.16.1, which tells you this is a maturing, iterating codebase, not abandonware.
Who should use Kubeflow Pipelines?+
Teams already running Kubeflow (or planning to) who need pipeline orchestration as part of that stack. ML platform teams building internal MLOps tooling on Kubernetes who want a Python-native SDK for defining pipelines with strong artifact/metadata tracking baked in.
Who should not use Kubeflow Pipelines?+
Small teams without dedicated platform/infra engineers. KFP requires standing up and maintaining Kubeflow (or at minimum KFP standalone) on Kubernetes — that's real operational burden, not a pip install.
What are the alternatives to Kubeflow Pipelines?+
Airflow — the incumbent general-purpose orchestrator; better ecosystem/integrations, worse for ML-specific artifact/metadata tracking. Argo Workflows — the engine KFP is built on; use directly if you don't need KFP's ML-specific abstractions and want less overhead.
How much does Kubeflow Pipelines cost?+
Fully open source, Apache-2.0. No paid tier, no vendor lock-in from the project itself. Your cost is the Kubernetes cluster and engineering time to run and maintain it — that's the real "price" of KFP, and it's not small.