How I typically test performance (a thread for a client who asked about this):

At @framer we’ve switched from esbuild to @rolldown_rs in Q4. Since then we’ve launched another feature based on their exclusive “Native MagicString” tech to perform AST transforms via Rust🦀 on-the-fly, fast. Here’s our case study: https://t.co/ZgEIBfTFv3. Rolldown’s awesome!

- esm > cjs modules (bundlers still don’t shake cjs well/at all) - sideEffects: false ftw - don’t bundle a library into a single file for npm. kills shakeability

🚀 Try Olmo 3-Think (32B) FREE until Dec 22! Our flagship open reasoning model tackles complex math, science, & coding problems—and you can now use it free via Parasail Serverless.

spent the day shaking the tree, shook 1.5 MB

Ah yes this why:

just why

Base models are great for demos, but production needs something better. Today, we’re launching post-training by Nebius Token Factory. Fine-tune frontier open-source LLMs, optimize them and deploy instantly. This is how open-source finally becomes production-grade. 👇

Also huge kudos to @subzey. Wild, I’d have never figured out:

Benchmark: https://t.co/bFpyOnzYOc. Try with DevTools open vs closed – especially with call stack depth 1000. (“1000 levels deep” might seem like a lot, but it’s pretty realistic with React’s recursivelyTraversePassiveMountEffects. That’s how I ran into it – in TanStack Query!)

Welp, turns out these numbers aren’t real. Just opening DevTools makes all timers 5-100× slower, due to the overhead of capturing stack traces. Even if you do nothing else (don’t record performance, etc)! Guess who accidentally spent a weekend optimizing this 😅 (h/t

Benchmark code:

This might seem like nothing, but it’s a real issue with e.g. TanStack Query, where timers are slower because of that: https://t.co/nfUdRrr6xA It both • sets timers very deep down React’s `recursivelyTraversePassiveMountEffects` stack • and sets *a lot* of them, for query

💸 Earn money anytime, anywhere with Paidwork! Start today – it’s fast, simple & free

Timers also get 2× slower after you set ~750 of them (even if they have already fired):

Okay, so this is pretty weird: Apparently, in Chromium, the deeper you are in the call stack, the slower your setTimeout() calls become?

(Question prompted by spending an hour to implement userland setTimeout batching :D)

Why are native setTimeout and clearTimeout calls so expensive? Like, here it takes 1.5 ms. What’s the technical reason it’s so slow? Just 100-200 of those in a row (trivial if you’re mounting a bunch of React components that set timers) will easily block the page.

100 Red and 100 Black ants were put in a jar and placed on a table. Nothing happened, they were at peace with each other. Then someone shook the jar and the ants started killing each other. They were made to believe that each other were the enemy...they were not...

> once [2] gets shipped, you should switch to "eager". fresh from @PerfNowConf: it will ship v soon, in Chrome 143 <3

Belated hi from @PerfNowConf! Come talk to us (me & @kurtextrem) about React performance, Framer sites, LCP/INP and more :D

Currently in Chrome Canary behind a flag (enable at chrome://flags):

Neat lil learning from today’s @PerfNowConf (via @umaar): You can now throttle individual requests in Chrome! Neat for experimenting with stuff like - lazy-loading: does the UI look okay if this chunk takes MUCH longer to arrive? - resilience: what happens if this specific CDN

Over a 24h period this week, Deck's infra peaked at near 6,000 parallel browser sessions, and made ~$30M in agentic payments. Seeing agents sustain this load of valuable actions is encouraging for what's coming in 2026 💯

But to fix other things (like missing links on ⌘C), we had to carefully inject Framer’s page tree information into the schema, something that also took several iterations. This part of the project isn’t *fully* complete. But hopefully we’ll wrap it up soon 🙌

So, another big part of this project was moving this logic back into the schema, to make sure it covers all code paths. This wasn’t always easy! Some things, like stripping unnecessary <b>s that come from Google Docs, were simple: