next.js

85c3f2e7 - Turbopack: bincode: Migrate TaskInput serialization to bincode (#86631)

Turbopack: bincode: Migrate TaskInput serialization to bincode (#86631) https://github.com/vercel/next.js/pull/86338 switches cell storage to use bincode, but it uses a compatibility shim for storing `TaskInput`s using `serde` + `pot`. This PR completes the migration by switching `TaskInput` to use `bincode` as well. ## Benchmark Summary `next build` is consistently faster across a variety of repositories when persistent caching is enabled, and the resulting cache size is consistently smaller.  `cargo build` runs only slightly faster, but there's potential for improving this further by removing unused `serde` macro derives. ## Benchmarking Notes Performance measurements (aside from cargo build) are taken using `next build`​. This should have proportional benefits for `next dev`​ too, but `next build`​ is more reproducible and easier to measure. `canary` refers to `a8f7b5467485647def4c0f4479a0a83bc4a3673a`. These numbers are taken with persistent caching enabled with an empty cache, so it's the worst-case scenario for the serialization codepath. We don't care nearly as much about the performance of deserialization because it's less common. ## `cargo build -p next-swc-napi` This makes builds _slightly_ faster. This branch doesn't clean up the serde derives yet, so there's a bunch of unused macro expansion here, so I think this number is actually pretty good. ``` ~/next.js $ hyperfine -w 1 -r 5 -p 'git co a8f7b5467485647def4c0f4479a0a83bc4a3673a && cargo clean' -p 'git co bgw/bincode-task-inputs && cargo clean' 'cargo build -p next-swc-napi' 'cargo build -p next-swc-napi' Benchmark 1: cargo build -p next-swc-napi Time (mean ± σ): 237.476 s ± 1.420 s [User: 2001.820 s, System: 84.957 s] Range (min … max): 235.294 s … 238.862 s 5 runs Benchmark 2: cargo build -p next-swc-napi Time (mean ± σ): 232.366 s ± 1.549 s [User: 1945.355 s, System: 83.940 s] Range (min … max): 230.550 s … 234.485 s 5 runs Summary cargo build -p next-swc-napi ran 1.02 ± 0.01 times faster than cargo build -p next-swc-napi ``` ## next-site Using `~/front/apps/next-site` on macos with a M2 Pro. Nothing else was running on the machine. ``` hyperfine -p 'rm -rf .next' -w 2 -r 20 'TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile' du -sh .next/cache/turbopack/ du -s .next/cache/turbopack/ ``` #### Canary ``` Benchmark 1: TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 14.999 s ± 0.258 s [User: 84.137 s, System: 8.683 s] Range (min … max): 14.400 s … 15.509 s 20 runs ``` ``` 855M .next/cache/turbopack/ 1751640 .next/cache/turbopack/ (Note: macos measures size in 512B blocks) ``` #### This Branch ``` Benchmark 1: TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 14.106 s ± 0.255 s [User: 79.976 s, System: 8.421 s] Range (min … max): 13.829 s … 14.939 s 20 runs ``` ``` 767M .next/cache/turbopack/ 1571072 .next/cache/turbopack/ (Note: macos measures size in 512B blocks) ``` ## vercel-site Using `~/front/apps/vercel-site` on macos with a M2 Pro. Nothing else was running on the machine. ``` hyperfine -p 'rm -rf .next' -w 2 -r 20 'TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile' du -sh .next/cache/turbopack/ du -s .next/cache/turbopack/ ``` #### Canary ``` Benchmark 1: TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 93.274 s ± 1.630 s [User: 671.120 s, System: 57.888 s] Range (min … max): 91.111 s … 96.881 s 10 runs ``` ``` 4.3G .next/cache/turbopack/ 9002568 .next/cache/turbopack/ (Note: macos measures size in 512B blocks) ``` #### This Branch ``` Benchmark 1: TURBOPACK_PERSISTENT_CACHE=1 TURBO_ENGINE_IGNORE_DIRTY=1 pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 88.694 s ± 0.963 s [User: 633.801 s, System: 54.576 s] Range (min … max): 87.357 s … 90.037 s 10 runs ``` ``` 3.7G .next/cache/turbopack/ 7788472 .next/cache/turbopack/ (Note: macos measures size in 512B blocks) ``` ## shadcn/ui Using a low-noise Linux machine: https://github.com/bgw/benchmark-scripts/ ``` diff --git a/apps/v4/next.config.mjs b/apps/v4/next.config.mjs index 7fa0f012..d1137d01 100644 --- a/apps/v4/next.config.mjs +++ b/apps/v4/next.config.mjs @@ -27,6 +27,7 @@ const nextConfig = { }, experimental: { turbopackFileSystemCacheForDev: true, + turbopackFileSystemCacheForBuild: true, }, redirects() { return [ ``` ``` cd ~/shadcn-ui/apps/v4 hyperfine -p 'rm -rf .next' -w 2 'pnpm next build --turbopack --experimental-build-mode=compile' du -sh .next/cache/turbopack/ du -s .next/cache/turbopack/ ``` #### Canary ``` Benchmark 1: pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 62.772 s ± 0.939 s [User: 163.153 s, System: 11.170 s] Range (min … max): 60.739 s … 64.220 s 10 runs ``` ``` 596M .next/cache/turbopack/ 609636 .next/cache/turbopack/ (Note: Linux measures size in 1024B blocks) ``` #### This Branch ``` Benchmark 1: pnpm next build --turbopack --experimental-build-mode=compile Time (mean ± σ): 59.017 s ± 1.014 s [User: 151.359 s, System: 9.978 s] Range (min … max): 57.124 s … 60.343 s 10 runs ``` ``` 491M .next/cache/turbopack/ 502240 .next/cache/turbopack/ (Note: Linux measures size in 1024B blocks) ```