next.js

08668326 - refactor: migrate to typed accessors and remove CachedDataItem adapter (#88397)

refactor: migrate to typed accessors and remove CachedDataItem adapter (#88397) Adopt the new generated accessor methods in turbo-tasks-backend. Remove the CachedDataItem enum and related support macros. This new approach is more ergonomic, memory efficient, and generates slightly smaller serialized payloads. From measuring vercel-site: ``` canary@e05510ab3bf91c02c7381e2c471363561ffa1198: $ hyperfine -p 'rm -rf .next' -w 2 -r 10 'pnpm next build --turbopack --experimental-build-mode=compile' Benchmark 1: pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 55.981 s ± 0.885 s [User: 386.683 s, System: 111.484 s] Range (min … max): 54.438 s … 57.723 s 10 runs ``` ``` 01-09-migrate-to-typed-accessors@d343fad5bf92af9c5271c62a24aedcf02c3d83d4 $ hyperfine -p 'rm -rf .next' -w 2 -r 10 'pnpm next build --turbopack --experimental-build-mode=compile' Benchmark 1: pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 54.897 s ± 0.496 s [User: 389.400 s, System: 127.353 s] Range (min … max): 54.298 s … 55.681 s 10 runs ``` so a small savings of a little over a second. Not too bad for mostly removing a bunch of `match` expressions that were required in the old model. similarly we should expect a small progression in disk size since we no longer explode into `Vec<CachedDataItem>` which induces some overhead from redundantly encoding enum descriminents and not being able to 'length prefix' collections. from a clean `.next` dir on vercel-site, and a production build with caching enabled, the cache size is before ``` $ du -h .next/cache 4.3G .next/cache/turbopack/v16.1.0-canary.0-518-ge05510ab3bf9 ``` after ``` $ du -h .next/cache 4.1G .next/cache/turbopack/v16.1.0-canary.0-546-gd343fad5bf92 ``` Which is expected, the main thing we are winning in terms of cache size is the CachedDataItem descriminent values, and in this build we were serializing about 260M items, so we save 1 byte for each of them. In some measurements i saw much more dramatic wins of >1gb, which i do not have a great explanation for.