Perf: various optimizations

[maartenbreddels]

Gives a ~25-30% performance boost for creating widgets (benchmarking against 1000 buttons) by

• not having comm a trait (no need for this)
• doing less for every instance (and caching it per class) (.keys and .to_json lookup)
• micro-optimization for state_get
  • not using unneeded _trait_to_json
  • also using the cached to_json (_trait_to_json_dict)
  • not checking not PY3 in loop
  • not checking drop_defaults in loop

Supersedes https://github.com/jupyter-widgets/ipywidgets/pull/3020

Benchmarks are done including this branch https://github.com/ipython/traitlets/pull/777

[filipre]

A somewhat wild guess but I think

not having comm a trait (no need for this)

might fix https://github.com/jupyter-widgets/ipywidgets/issues/3448. At least it would make debugging the issue a bit easier. What are the next steps for this PR?

[danielhollas]

Hi :wave:

Just wanted to signal our interest in this PR. :blush: I've tested this branch and by my (very rough) benchmark, this would save us around 300 ms startup time for our app. :open_mouth: (the linked traitlets PR already saved us 600ms, thanks so much @maartenbreddels for your work!)

Not sure if I can be of any help here besides testing. I've tried to merge with the current 7.x branch and there were only two minor conflicts that should be easily resolvable. I did not try to put this on top of 8.x branch.

[maartenbreddels]

Just wanted to signal our interest in this PR. 😊

Thank you :)

I'm experimenting with this in solara, adding it to ipyvue at runtime https://github.com/widgetti/solara/pull/389

I think the not having comm a trait feature might be something more for ipywidget9, as it might be a breaking change.