Storage and databases
Relational, vector, object, shared-file, model, and backup storage for platform services and tenants.
Foundation uses several persistence patterns because platform state, artifacts, model weights, and backups have different consistency and access requirements.
Relational state
CloudNativePG operates PostgreSQL clusters for transactional service state. PgAdmin gives authorized operators a management and inspection interface. PostgreSQL is used for metadata, configuration, workflow state, authorization state, and other structured records.
Tenant databases live on the same shared CloudNativePG cluster. When PostgreSQL is selected for a tenant, the tenant operator creates the tenant application database, which workloads in the tenant namespace use as the <slug>_app role. A tenant with private MLflow also gets the <slug>_mlflow metadata database. Each database has its own owner role, and tenant roles reach only their own databases. See Tenant application database.
Vector storage
Milvus stores the vectors behind retrieval-augmented generation and knowledge-base collections. Its segments and indexes live in object storage. When a tenant's Milvus capability is enabled, the tenant operator gives the tenant its own Milvus database and machine identity, whose permissions apply to that database only.
Object and shared-file storage
By default, Rook-Ceph provides S3-compatible object storage through RGW. Platform services such as MLflow, Milvus, Kubeflow Pipelines, Argo Workflows, Argo Events, Loki, Mimir, Tempo, Lakekeeper warehouses, and the BSQAI API store artifacts, data, and backups there.
Each tenant with storage gets its own bucket, <slug>-data. It holds the tenant's library files, its private MLflow artifacts, and, where the site enables the agent sandbox, its sandbox workspaces. The Global Organization's sandbox workspaces use the sandbox-workspaces bucket.
Persistent volumes use the site's [common] rwo_storageclass and rwx_storageclass. Rook-Ceph can also back those classes, with block volumes through RBD and shared filesystems through CephFS. In platform 2.0.0 and later, Rook-Ceph runs these modes only when a site storage class names them. A site that uses its infrastructure provider's storage classes runs Rook-Ceph as object storage only. See Ceph storage modes.
Deployments can instead configure an external S3-compatible object store with [object_storage]. The configuration includes private and public endpoints, credentials, TLS CA material, and capability flags so components do not assume that every provider supports Rook-specific APIs. External buckets must allow portal-origin GET and HEAD requests for browser-facing presigned downloads. The storage owner can manage that CORS document directly or explicitly delegate it to the platform with manage_bucket_cors = true.
See External S3 object storage for the bucket inventory, credential variables, configuration example, validation, and migration procedure.
Model storage
Model Installer can load models from Hugging Face, PVCs, S3-compatible storage, or the MLflow model registry. A tenant's models load from its own bucket or its private MLflow. KServe runtime profiles can stream Safetensors from the platform object store. An optional [model_storage] section defines a separate shared S3 repository for model artifacts when models should not use the primary platform object store.
Database backups
CloudNativePG backups use the Barman Cloud plugin and an ObjectStore resource. The default object destination is the cnpg-backups bucket, with retention controlled through platform configuration.
Why the separation matters
- PostgreSQL provides transactions and queries.
- Milvus provides vector search for retrieval.
- Object storage holds large artifacts, backups, and data products.
- Block and shared-file storage, from Rook-Ceph or the site's own storage classes, support stateful Kubernetes workloads.
- Dedicated model storage allows large weights to be shared without coupling them to application buckets.