Interactive TensorBoard, TensorFlow's Visualization Toolkit, session
This Tensorboard OnDemand app runs through an authenticating reverse proxy.
The app requires ood-util and ood-initializers to work.
About
TensorBoard normally starts on a defined port and requires no authentication. Anyone that can access the compute node by ssh can then also access TensorBoard. TensorBoard uses Werkzeug, which supports Unix sockets, but this functionality is not made available by TensorBoard as passing a Unix socket as host makes it crash when trying to parse the URL for printing it to the user.
After patching tensorboard/program.py with template/tensorboard.patch, TensorBoard on an Unix socket is fully functional and can be launched as tensorboard --host=unix:///path/to/socket.
We can then use a Nginx reverse proxy to the Unix socket that checks that the authentication cookie is correct.
This way we can have a fully authenticated TensorBoard available through OOD.