Where Does Kubeflow Store Pipeline Data?
Kubeflow keeps no local files: metadata in a MySQL database and artifacts in an object store, both on cluster volumes. What lives where.
Last updated
Kubeflow stores nothing on your laptop. Pipeline metadata (experiments, jobs, runs, scalar metrics) lives in a MySQL database inside the cluster. Artifacts (pipeline packages, views, time series metrics) live in an object store bucket on cluster persistent volumes.
There is no config file to copy for a backup. Protection means volume snapshots plus database dumps, or a second cluster with replication. Local kubectl output only ever shows pointers, never the data itself. Upgrades keep these stores in place, which is why a fresh install can adopt an old cluster's data.
Where Kubeflow stores this, by platform
mlpipeline MySQL database and object store bucket inside the cluster (see notes)
Metadata in the mlpipeline MySQL database, artifacts in the mlpipeline object store bucket (MinIO historically, SeaweedFS default since KFP 2.15). Both sit on Kubernetes PersistentVolumes. Configure the artifact path per pipeline with defaultPipelineRoot in the kfp-launcher ConfigMap.
Frequently asked questions
How do I inspect Kubeflow data directly?
Port forward the services and query them. The mlpipeline MySQL database holds runs and experiments, while the object store bucket holds pipeline files and artifacts. kubectl exec plus the MinIO or SeaweedFS console covers both.
Is the artifact store MinIO or SeaweedFS?
Yes. KFP 2.15 switched the default object store deployment from MinIO to SeaweedFS, with MinIO still supported. Older installs use MinIO buckets, so check which deployment your cluster runs before following a guide.
Notice an outdated path? Let us know.