TensorBoard plugin blocks TensorBoard UI from fully loading

[psybuzz]

Repro:

• In a fresh Python virtualenv, pip install tensorboard torch-tb-profiler
• Open TensorBoard (http://localhost:6006 by default)
• The UI under the header is blank, but the loading spinner rotates
  
• In the DevTools network panel, the '/plugins_listing' request is stalled
• Eventually showing an error
  

On my machine, I tested this with

• tensorboard==2.5.0
• torch-tb-profiler==0.2.0

From a brief read, torch-tb-plugin's is_active() method seemingly issues blocking call to wait for data:

kineto/tb_plugin/torch_tb_profiler/plugin.py

Lines 69 to 74 in be239c0

	def is_active(self):
	"""Returns whether there is relevant data for the plugin to process.
	"""
	self._is_active_initialized_event.wait()
	return self._is_active

On the TensorBoard side, there have been a few reports from users who find TensorBoard responsive again after uninstalling torch-tb-plugin: tensorflow/tensorboard#5076, tensorflow/tensorboard#5111

TensorBoard's other builtin plugins are implemented with is_actives that either return quickly or are non-blocking. Would it be possible for torch-tb-plugin to do something similar?

[EricWiener]

I'm also running into this issue with a docker image w/ tensorflow 2.5.0, tensorboard 2.5.0, and torch_tb_profiler 0.2.0.

[guotuofeng]

This is a known bug, we will fix it