Data model
Our data model is inspired by the MLflow machine learning lifecycle and simplified to align with our general Workers’ developer experience.
The three fundamental concepts of our model are projects, models, and the catalog.
Projects
The project has a name, a machine-learning runtime, and an ID.
The runtime has to be one of our supported engines. Refer to Constellation Runtimes for the complete list. The ID is what binds the project to the Worker.
The name can only have alphanumeric, minus (-) and underscore (_) characters (/^[a-zA-Z0-9-_]+$/). The name helps with organization and is used for certain operations in Wrangler and the APIs.
You can have as many projects as you want under your account.
Models
The models are user-uploaded files that are attached to a specific project. The model has a name, a description, and an ID. A model needs to be compatible with the machine-learning runtime defined in the project it belongs to. For example, you should only upload the model.onnx file to a project configured for ONNX runtime.
The name can only have alphanumeric, minus (-) and underscore (_) characters (/^[a-zA-Z0-9-_]+$/).
You can have as many models per project as you want.
If you delete a project, you also delete all its associated models.
Currently, during the private beta, we only support models that are smaller than 10 MiB.
Catalog
You may not want to train models or upload models you have not tested. Cloudflare will maintain a list of verified models that are known to work with the Constellation APIs without extra configuration.
For each machine-learning runtime we support, developers can search for ready-to-use permanent models in our catalog and use them for some of the most popular tasks without additional configurations or file uploads.
Like user-uploaded models, the models in the catalog have a name, an ID, and a description and are associated with a parent catalog project that defines the runtime.