Google-internal Colab cannot load scalar cards in tb-nightly

[stephanwlee]

Scalars recently have migrated to use POST request for optimizing the data loading--for a given tag, we make ~1 request with multiple runs in the payload. This optimization seem to have broken the Colab integration as the Colab network proxy does not relay POST requests.

Googlers, internal reference: b/168798361

[wchargin]

Hmm; I can’t reproduce this:

!pip uninstall -y tensorboard && pip install tb-nightly
import tensorflow as tf
with tf.summary.create_file_writer("logs").as_default():
  tf.summary.scalar("loss", 0.1, step=0)
%load_ext tensorboard
%tensorboard --logdir logs

Inspecting network traffic shows that the scalars_multirun request is
being sent as POST. For me, it returns the right data and renders fine.
The Colab tunnel relays POSTs as of http://b/126387106, which went
live by default in April 2019.

[wchargin]

It looks like this issue only affects Google-internal users. See my
comments on http://b/168798361.

[wchargin]

We decided to fix this anyway: see #4191.