Model release cycle

[ucbrise]

Model release cycle

[dcrankshaw]

Yes this is explicitly supported in Clipper.

Clipper is designed to support frequently retrained models with minimal overhead. A model deployed in Clipper is composed of two elements, the model container and the model data. The model container is the code (and any dependencies) needed to evaluate the model. The model data is the state (generally the trained model's parameters in some serialized format) that is loaded when a container is initialized. For frequently retrained models, the code often doesn't change, just the parameters from being trained on new data. This means that the updated model can be deployed without needing to write any code or modify the container at all by deploying a new version of the model with new model data. This can all be done programmatically.