refactor(core): Use iterative approach in recompute dependents

[dmaskasky]

Summary

Replaces the recursive topo search in recomputeDependents with an iterative one. An iterative approach theoretically is faster and does not suffer from stack overflow issues. This makes it possible to process very large atom graphs.

Benchmark

• Recursive: ~200ms
• Iterative: ~ 250ms
  TODO: Currently the new iterative approach is slower than the recursive one.

Check List

• pnpm run prettier for formatting code and docs

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
jotai	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Nov 17, 2024 11:00am

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.

[github-actions]

Preview in LiveCodes

Latest commit: d570004
Last updated: Nov 17, 2024 10:59am (UTC)

Playground	Link
React demo	https://livecodes.io?x=id/8ZPZWPCBW

---

See documentations for usage instructions.

[dai-shi]

This looks like a great improvement.

[dai-shi]

LGTM

[cltnschlosser]

Can you share the benchmark that you used to get the times in the PR description? This slowdown isn’t going to work for us, but I think I see how to optimize the iterative case, I want to prove it with the same benchmark though.

[dmaskasky]

Oh, would you be able to provide these benchmarks? I only used vitest test duration using stack limit size nodes (~12k) in a chained dependency configuration.

I'd love to hear your ideas on how to make this faster. The original recursive solution suffers from stack overflow issues, so I'm keen on keeping an iterative approach. I tried both DFS and BFS and found they both have the same number of iterations but the DFS was more memory efficient.

[dmaskasky]

Oh, I found a typo in the PR description. The iterative DFS solution in around 250 ms in my testing, not 350 ms.

[cltnschlosser]

We should be able to shuffle some things around to remove one of the visiting or visited sets.

[dmaskasky]

I'm not sure how to improve on the post-order DFS topological sort solution we have currently. I tried implementing Kahn's, but there's too much administrative work required to beat the current solution.

#2833