MLflow

The model lifecycle, registry, and GenAI observability service used by AI.

Agentic Friendly

Component Category

Model registry, ML lifecycle, and GenAI observability

Component Description

MLflow is the model lifecycle service used in AI for experiment tracking, model registration, controlled model version management, and GenAI observability. BullSequana AI deploys MLflow 3.17.2.

The Global organization uses the installation's MLflow. A customer organization can have its own private MLflow, which the BSQAI Tenant Operator runs for the tenant with its own tracking database and an artifact prefix in the tenant's bucket. Only the BSQAI API reaches either instance. Experiments, prompts, and registered models belong to the person who created them and are shared with grants; their ownership can be transferred. See MLflow follows the active organization.

Why It Is Used

In BullSequana AI, MLflow gives teams a structured way to manage model artifacts and lifecycle steps before models are exposed to applications and inference services. The BSQAI API also sends GenAI traces to an experiment in the installation's MLflow through an isolated OpenTelemetry provider.

Learn More

Artifact downloads

The MLflow server presigns artifact downloads with the object storage endpoint that clients outside the cluster reach: the public S3 hostname of the platform's Rook-Ceph storage, or the configured endpoint of an external S3 service. MLflow is internal-only: tenants reach it only through the BSQAI API, which downloads artifacts from that endpoint and streams them to the Portal and API callers. Platform jobs and monitoring also connect to it inside the cluster.

On the public Rook-Ceph S3 route, a transfer has no total time limit. The route closes a stream after five minutes without activity. Client timeouts and the expiry of the presigned URL still apply.

Interacts With

  • AI Web Portal for user-facing lifecycle interactions, including the Studio model repository and experiments browser
  • BSQAI API for backend model-aware workflows and AI observability traces
  • Model Installer, part of the BSQAI API, for deploying registered LLM and predictive model versions to KServe in the organization's namespace
  • BSQAI Tenant Operator for provisioning each customer organization's private MLflow
  • OpenTelemetry for GenAI trace export
  • Rook-Ceph or external S3-compatible storage for artifact storage
  • PostgreSQL for persistent service state

On this page