[not for merge] experiment: evaluate various strategies to improve performance

[dmaskasky]

Related Bug Reports or Discussions

#3275

Summary

Does the building block weak map negatively impact performance? This PR is an experiment to see if moving back to Symbol helps.

Check List

• pnpm run fix for formatting and linting code and docs

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
jotai	Ready	Preview, Comment	Mar 30, 2026 2:23am

[pkg-pr-new]

More templates

• hacker_news
• hello
• mega-form
• starter
• text_length
• todos
• todos_with_atomFamily

npm i https://pkg.pr.new/jotai@3276

commit: 7a19d66

[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: 7a19d66
Last updated: Mar 30, 2026 2:22am (UTC)

Playground	Link
React demo	https://livecodes.io?x=id/5DMUJNTQU

---

See documentations for usage instructions.

[aio39]

I ran some benchmarks to see how this change performs in practice. While the Symbol approach does show a nice improvement in atom creation (-25%), the overall impact on most hot paths appears to be modest (~1–4%). I hope this data is helpful for evaluating the direction.

Benchmark Results: WeakMap → Symbol (PR #3276)

Ran the same benchmark suite from #3275 against three versions: v2.10.0 (baseline), v2.19.0 (current), and this PR (built from source with the Symbol change applied on top of v2.19.0).

• 10 iterations per version, randomized sequential order
• 3 warmup runs + global.gc() + 1 timed measurement per scenario
• Node.js with --expose-gc, NODE_ENV=production
• Benchmarked with Claude Code AI assistance

Results (median of 10 runs, ms — lower is better)

Scenario	v2.10.0	v2.19.0	PR #3276	PR vs v2.19.0	PR vs v2.10.0
Atom Creation (10k)	0.747	0.812	0.607	-25.3%	-18.8%
Primitive R/W (10k)	10.445	71.192	68.814	-3.3%	+558.8%
Derived Chain (depth=100)	1.850	1.656	1.594	-3.7%	-13.8%
Wide Fan-Out (1→1000)	7.265	16.754	16.557	-1.2%	+127.9%
Diamond (100 mid)	0.796	1.934	1.997	+3.3%	+150.8%
Subscription Churn (1k)	1.915	6.800	6.783	-0.3%	+254.1%
Computed Read (10k)	2.331	13.393	12.942	-3.4%	+455.1%
selectAtom (10k)	26.808	146.997	143.109	-2.6%	+433.8%

Summary

• Atom creation sees a clear -25.3% improvement — the most impactful gain from this PR.
• Most other scenarios show 1–4% improvement over v2.19.0, consistent with reducing WeakMap.get() overhead in the hot path.
• Diamond (+3.3%) is within noise range.
• However, the gap to v2.10.0 remains large (100–560%) across most scenarios, indicating that the WeakMap → Symbol swap alone does not address the core regression introduced around v2.14.0. The frequent getInternalBuildingBlocks() calls and the overall buildingBlocks indirection architecture appear to be the larger bottleneck.

Ref: #3275