Rethink the model server and store

[hnarayanan]

The existing system using TorchServe is quite basic in terms of model management. A much more powerful system would involve:

• MLflow: For experiment management and model versioning
  • TorchServe: The actual core serving can still be handled by TorchServe plugging into MLflow
• Airflow: For sequencing of scripting around training

The core of this is explained by another project in the class. They also have a GitHub repository that might be useful.