perf(routing): cache app-route-graph directory reads — 5000-route build −32%

[hyf0]

No dependencies — reviewable and mergeable on its own. Touches only routing/app-route-graph.ts.

What

Memoizes the raw fs.readdirSync(dir, { withFileTypes: true }) calls in the App Router graph scan, via a per-scan WeakMap<ValidFileMatcher, Map<string, Dirent[]>> (the same pattern as the existing findFileProbeCache / findSlotSubPagesCache). discoverParallelSlots and findSlotRootPage read through it.

Why

discoverInheritedParallelSlots reads every ancestor directory of each route to look for @slot dirs; because routes share ancestors, a directory with N sibling routes is re-read once per descendant route — super-linear in route-directory width. findFileProbeCache memoizes findFile, but not these raw readdirSync/existsSync calls. The graph is scanned both during build and at dev-server startup.

Measured

Before/after this cache on the generated stress app (#2388 adds the 5000-route scenario), local wall-clock. The effect is super-linear, so it grows with scale:

Scale	Metric	Baseline	With cache	Change
5,000 routes	Production build	~30 s	~20 s	−32%
5,000 routes	Dev-server cold start	~29 s	~10 s	−67%
10,000 routes	Production build	~86 s	~42 s	−52%
10,000 routes	Dev-server cold start	~101 s	~17 s	−83%

readDirEntriesCached returns exactly the entries fs.readdirSync would for each directory, so the discovered route/file set — and the module transforms that follow — are unchanged (verified by an identical module-transform count). Dev cold start improves because the route scan runs synchronously at every dev (re)start, independent of the .vite cache — so this isn't only a cold-cache effect.

At scale vs Next.js (Turbopack)

With this cache in place, vinext's build-time lead over Next.js (Turbopack) widens sharply with route count — same machine, same generated app, clean production builds (median of 3, local wall-clock):

Scale	vinext	Next.js (Turbopack)	vinext faster
33 routes	2.6 s	5.1 s	1.95×
5,033 routes	20.2 s	106.0 s	5.26×

Marginal cost per added route ≈ 3.5 ms (vinext) vs 20 ms (Turbopack). Keeping the route-graph scan near-linear (this PR) is part of why vinext holds its lead as apps grow.

Memory safety

The WeakMap is keyed on the per-scan matcher clone (like findSlotSubPagesCache), so entries are released when the scan's matcher is GC'd — nothing accumulates across rebuilds/restarts in a long-lived dev server (a module-level Map would leak here; the WeakMap-per-scan-matcher does not).

Correctness

• readDirEntriesCached caches [] only for a missing directory (ENOENT/ENOTDIR) — exactly replacing the prior fs.existsSync(dir) guard — and rethrows any real fault (EACCES, EMFILE/ENFILE, …), matching the original unguarded readdirSync. A genuine FS error still fails the build loudly rather than silently caching an empty listing and dropping slots. The cache returns exactly the entries readdirSync would, so the discovered file set and resulting transforms are identical (same module-transform count).
• Route/slot suites pass — app-route-graph, slot, app-optimistic-routing, routing, route-sorting, app-rsc-route-matching, hybrid-route-priority, app-page-route-wiring, pages-router, intercepting-routes-build, app-browser-interception-context: 662 tests. vp lint clean.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@vinext/cloudflare@2389

npm i https://pkg.pr.new/vinext@2389

commit: d64ac0b

[github-actions]

Performance benchmarks

Compared d64ac0b against base da086c0 using alternating same-runner rounds. Next.js was unchanged and skipped.

0 improved · 0 regressed · 6 within ±1.5%

Scenario	Framework	Baseline	Current	Change
Client bundle size (gzip)	vinext	122.0 KB	122.0 KB	⚫ +0.0%
Client entry size (gzip)	vinext	117.0 KB	117.0 KB	⚫ +0.0%
Dev server cold start	vinext	2.46 s	2.48 s	⚫ +0.8%
Production build time	vinext	2.94 s	2.94 s	⚫ -0.2%
RSC entry closure size (gzip)	vinext	94.3 KB	94.3 KB	⚫ +0.0%
Server bundle size (gzip)	vinext	160.0 KB	160.0 KB	⚫ +0.0%

View detailed results and traces

_{🟢 improvement · 🔴 regression · ⚫ change below 1.5% · paired base/head}