Browser access
The server refuses anonymous access. Team members sign in with their own Trackshund identity, which grants access to their team’s server.
Security boundaries for a private MLflow server
This page explains what Trackshund separates between customers, how people and scripts get access, and what happens to data after a server is deleted.
What Trackshund separates
| Boundary | What each customer gets | Why it matters |
|---|---|---|
| Service | A separate MLflow container and database | One team does not query another team’s metadata store |
| Network | A separate internal network and exact hostname route | An unknown hostname does not select a customer server |
| Artifacts | A restricted namespace and credential | The server can reach its own artifact objects, not a neighbour’s |
| Resources | CPU, memory, process and log ceilings | One workload has defined limits on a shared host |
Individual access
Scripts and CI
The MLflow server issues and revokes its own access tokens. Trackshund does not put a tenant’s script token in the control-plane database.
Operator access
The control plane deliberately has no account inside a customer’s MLflow server. Operational checks verify that anonymous access is refused; they do not read experiments.
Backups
The metadata and access-control databases are copied nightly using SQLite’s online backup mechanism. Seven nightly generations are kept outside the customer-accessible artifact namespace, so a customer token cannot delete the backup beside the original.
Deletion
Deleting a server immediately removes its reachable address, workload, and credentials. Trackshund retains the underlying data for a 30-day operator safety window, then erases it.
Limits of this security model
Trackshund has not published an availability commitment or claimed a compliance certification. Physical hosts and their kernels are shared. The isolation model assumes customers cannot upload or run arbitrary code inside the managed service; adding that capability would require revisiting the model.
If these boundaries fit your requirements, deploy a private MLflow server.
One private MLflow instance for your team, at one flat monthly price.