Scaling Stylo's rayon-based parallelism beyond 64 cores required revisiting work-stealing queue contention. We saw speedup plateau around 48 cores on Graviton4 until we switched from a global queue to per-NUMA-node queues.
V8 popularized the hidden-class approach, but SpiderMonkey's inline-cache polymorphism has aged better for modern JS patterns. We measured a 12% speedup on real-world React benchmarks after the TI-to-IC refactor landed.
A surprise result from flamegraph profiling: 180ms of cold-start was spent registering URL schemes the fetcher never used. Removing them cut wall-clock by 38% on static-small workloads.
We hit a rustls init race where two crawl_each calls both tried to install the default provider. The fix was trivial once identified, but the symptoms were anything but obvious.