Mechanism for persistent user settings

[manivaradarajan]

As mentioned in #893 (e.g., in this comment, https://github.com/tensorflow/tensorboard/issues/893#issuecomment-668712228), we need a mechanism for persisting user settings / preferences.

I'm capturing some items / ideas here that should be put into a RFC / requirements doc.

• This needs to work both in regular TensorBoard and in TensorBoard.dev, etc.
• Tiered preference system:
• Global preferences (per user): example: smoothing
• Experiment specific preference (when viewing this experiment, change the run color to #aaa)
• We may want to make smoothing weight, ignore outlier experiment specific OR global (i.e., it is unclear what is more correct)
• Shareable UI configuration (so your colleague sees exactly the same UI when you share a URL)

Based on above considerations, schema/mechanism of the backend will change (simple KV pair or relational structure)

[stephanwlee]

We have made some progress in this bug. Now we use LocalStorage (not a panacea--if you want per-experiment settings like run selection, LocalStorage cannot scale to N-number of experiments; we can consider having some kind of LRU scheme but there could be a one logdir with many runs that can exhaust the size limit of LocalStorage) for persisting global settings like soothing, ignore outlier, reload, etc...

Until we have more bulletproof solution for "Experiment specific preference", we can keep this issue opened.