|
History
|
||
|---|---|---|
| .. | ||
| link-performance | ||
| react | ||
| scenarios | ||
| solid | ||
| vue | ||
| jsdom.ts | ||
| package.json | ||
| README.md | ||
| setup-helpers.ts | ||
| tsconfig.json | ||
| vitest.config.ts | ||
| vitest.react.config.ts | ||
| vitest.setup.ts | ||
| vitest.solid.config.ts | ||
| vitest.vue.config.ts | ||
Client Navigation Benchmarks
Cross-framework client-side CPU benchmarks for:
@tanstack/react-router@tanstack/solid-router@tanstack/vue-router
The benchmarks run in jsdom against production builds of real apps, and are tracked in CI by CodSpeed (simulation mode).
Scope: these benchmarks cover the client router. The hydration scenarios also cover DOM hydration and the Router state-restoration path used by Start. Start-specific entry-point initialization, streamed payload arrival and client server-function calls are not included. Server request work is covered by
benchmarks/ssr.
Layout
react/,solid/,vue/- baseline benchmark (mixed navigation loop) + Vitest configvitest.react.config.ts,vitest.solid.config.ts,vitest.vue.config.ts- per-framework aggregate configs that run the baseline first, then scenario projectsscenarios/harness.ts- shared scenario runner (mount, link-click steps,onRenderedsynchronization)scenarios/<scenario>/shared.ts- framework-agnostic scenario definition (workload data, step sequence, assertions, bench options)scenarios/<scenario>/<framework>/- isolated scenario apps
Scenario app layout:
scenarios/<scenario>/<framework>/
vite.config.ts
speed.bench.ts
speed.flame.ts
setup.ts
project.json
tsconfig.json
src/
main.tsx
routeTree.gen.ts
routes/
Navigation scenario apps use file-based routing (@tanstack/router-plugin) with a
generated routeTree.gen.ts, like a regular user app. Each scenario uses one
app per framework instead of sharing routes in the baseline app. This keeps
route-tree size and router options isolated so one scenario cannot shift
another scenario's numbers. The existing baseline apps and bench names stay
stable for CodSpeed continuity.
Scenario Responsibilities
Each scenario isolates one client-side responsibility so benchmark changes can be attributed to a specific feature area.
| Scenario | Client-side responsibility |
|---|---|
react/, solid/, vue/ baseline apps |
Mixed navigation loop: path params, search params, route context, and useParams/useSearch/useLoaderData selector subscriptions. |
async-pipeline |
The router's async pipeline via counted 0ms timer hops: async loaders (transition-held navigation), async beforeLoad context, and parallel nested async loaders. Component-level Await/Suspense is excluded: React 19 throttles Suspense reveals by ~300ms wall-clock, which is inherently non-deterministic to benchmark. |
control-flow |
Loader-thrown redirect (including a 2-hop chain), notFound() with notFoundComponent, loader errors with errorComponent, and boundary reset on recovery navigation. |
head |
HeadContent per-navigation work: nested route head() evaluation, title/meta/link dedupe across matches, and head tag DOM updates during navigation. |
history |
History push/replace/back/forward traversal, location masking, registered-but-never-blocking useBlocker, and useCanGoBack/useLocation subscriptions. |
hydration |
Initial DOM hydration: execute the SSR payload, restore beforeLoad context and loader data, and hydrate 192 ordinary and eight hash-sensitive Links through their follow-up effects in React and Solid. |
link-churn |
Links that depart, stay and mount: 30 /hub layout Links stay while 100 /hub/list leaf Links leave (alone, or with the layout to /other) and remount on return; a search-only change rebuilds updater and to="." Links. Unlike links, whose root-layout Links never depart. |
links |
Per-navigation cost of ~200 mounted <Link>s: link prop building, active-state recompute across activeOptions variants, activeProps swaps, and useMatchRoute probes (the MatchRoute component is avoided: vue-router's implementation leaks one subscription per render). |
loaders |
Client loader dispatch: always-stale re-runs (staleTime: 0), cached revisits (re-run once per lap by the invalidate step), loaderDeps-keyed caching, router.invalidate(), and useLoaderData selectors. |
mount |
Cold start: createRouter (route-tree processing) + first render + initial router.load() + unmount, with a fresh router per mount. |
nested-params |
Deep nesting (8 dynamic levels): per-level params.parse/stringify, beforeLoad context accumulation across matches, and per-level useParams/useRouteContext subscriptions. Param values include characters requiring percent-encoding (as do route-tree-scale's), so segment encode/decode paths run on every navigation. |
preload |
Intent preloading from hover events, programmatic router.preloadRoute, deterministic preload cache behavior (defaultPreloadStaleTime: 0), and commit-time cache maintenance. |
rewrites |
Composed client-side location rewrites: router basepath plus a locale input/output rewrite pair, running on every href build and location parse (the client analog of the SSR rewrites scenario). |
route-tree-scale |
Route matching and link-target resolution on a wide (~40 route) tree mixing static, dynamic, prefixed-param, splat, pathless-layout, and route-group paths, with autoCodeSplitting enabled so navigations also resolve lazy route chunks. |
search-params |
validateSearch execution, search middlewares (retainSearchParams/stripSearchParams), functional search updaters, structural sharing, and useSearch selector subscriptions. |
Conventions
The hydration scenario has the separate lifecycle described below. The other scenarios follow these navigation/mount conventions:
- Apps are built with
NODE_ENV=production(minify: false) intodist/app.js; benches import the built bundle, so production package builds and production JSX output are measured, not dev transforms. - Scenarios behave like a real user app: navigation happens through
<Link>clicks dispatched on real anchor elements (unless a scenario specifically measures the imperative API), the router uses the default browser history, andscrollRestorationis enabled. - Each benchmark iteration advances a fixed, circular sequence of steps; every step awaits the router's
onRenderedevent, so render work is included and steps cannot overlap. No two consecutive steps may target the same location, and the sequence ends back on the initial route. - Setup runs one warm-up lap through the sequence and asserts each step's observable output (e.g.
document.title), so a scenario that silently stops doing its work fails instead of reporting a fast time. - Apps are mounted and disposed outside timing by
benchLifecycle(benchmarks/cpu-simulation.ts): CodSpeed measures a single invocation, so every invocation gets a freshly mounted app; plainvitest benchkeeps one app per warmup/run phase. Bench functions must beasyncfunctions, because CodSpeed's Vitest 5 provider awaits onlyAsyncFunctions. - Determinism: no wall-clock timers (async work is resolved promises or counted
setTimeout(0)hops),staleTime/gcTimeonly ever0or effectively-infinite, noMath.random/Date.nowin the measured loop. Scroll restoration usesgetScrollRestorationKey: (location) => location.pathnameso its cache stays bounded (the default random per-entry key grows one cache entry per push navigation). Themountscenario opts out ofscrollRestorationentirely and destroys its history on unmount — both register page-lifetime globals that would leak across its mount/unmount loop. - Push-only step laps grow jsdom's session-history entry list over a run (jsdom never truncates entries behind the current index). Per-push cost stays O(1) so timings are unaffected; only the
historyscenario needs — and has — a lap that is depth-stationary. - Calibration: pick per-iteration step counts so a bench's
vitest benchrun stays roughly between 8 and 30 seconds (long enough to average out, short enough for CI).
Run
Run all benchmarks through Nx so dependency builds are part of the graph:
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:perf --outputStyle=stream --skipRemoteCache
Run framework-specific benchmarks (baseline + all scenarios):
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:perf:react --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:perf:solid --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:perf:vue --outputStyle=stream --skipRemoteCache
Run a single scenario app manually (after building it through Nx):
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav-<scenario>-<framework>:build:client --outputStyle=stream --skipRemoteCache
cd benchmarks/client-nav && NODE_ENV=production vitest bench --mode benchmark --config ./scenarios/<scenario>/<framework>/vite.config.ts
Run framework-specific flame benchmarks (10 second loop, profiled with @platformatic/flame, forced to NODE_ENV=production):
# baseline
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:flame:react --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:flame:solid --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:flame:vue --outputStyle=stream --skipRemoteCache
# scenarios
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav-<scenario>-<framework>:test:flame --outputStyle=stream --skipRemoteCache
Typecheck benchmark sources (baseline + scenarios):
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav:test:types --outputStyle=stream --skipRemoteCache
Hydration
scenarios/hydration/{react,solid} provide the same initial-hydration
workload: 192 ordinary Links, eight hash-sensitive Links, and three matched
routes with three beforeLoad contexts and two loader results. The server URL
has no fragment; the client URL has #details, matching half of the hash-sensitive Links.
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav-hydration-react:test:perf --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav-hydration-react:test:unit --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/client-nav-hydration-react:test:types:client --outputStyle=stream --skipRemoteCache
Replace react with solid to run the other adapter.
The build generates static HTML and its real Router SSR bootstrap scripts once
with createRequestHandler and the adapter's streaming renderer. Generation
fully consumes the response and runs outside the CodSpeed action. The measured
worker only reads these artifacts; it does not import a server renderer or start a server.
Each invocation uses a fresh jsdom window and a fresh evaluation of the complete production client bundle in that window's realm. HTML parsing, bundle evaluation, and the initial document lifecycle finish in untimed setup. This resets module state as well as the DOM, avoiding cached hydration promises and cross-realm payload objects. The bundle and bootstrap scripts are compiled once, so this measures a fresh application with warm code, not JavaScript download/parse cost.
The timed region executes the serialized payload, creates the router and its
small code-based route tree, calls the public client hydrate API, and hydrates
the existing DOM with the framework's native renderer:
- React: a separate completion component signals from its post-hydration
effect, after the same snapshot transition as hash-sensitive Links. The harness
awaits that signal and two idle React scheduler turns, then checks the expected
active links. The window's
performance.now()is frozen, so React's scheduler yields only to paint after commits, not after 5 ms of wall time, which CPU instrumentation stretches by a runner-dependent factor. Each sample therefore uses the same number of scheduler tasks.Date.now()is unchanged. A 60-second watchdog fails a sample that never settles. - Solid: execute the native hydration bootstrap and retain the server's component/key hierarchy. Wait for mount, the router's rendered event, active-link effects, and two idle turns. DOM identity assertions include every workload element so template fallback cannot silently replace server nodes.
Solid bounds settlement to 100 turns. Ending at the hydration call or the
first mount would miss post-hydration Link updates.
Timer turns use setImmediate; scrolling is a no-op
because this is CPU simulation rather than browser layout/paint measurement.
Untimed assertions verify restored contexts and every loader row, zero client
beforeLoad/loader calls, expected hrefs/active state, original DOM-node identity,
working event handlers, and absence of hydration errors. Root unmount, pending
task cancellation, and window disposal also run outside measurement. Diagnostic
tests count Link renders or reactive evaluations and check they stop before the
measured region ends; counting is disabled in the timed workload.
The bench's own beforeEach/afterEach hooks prepare and dispose every
sample outside timing, in both CodSpeed and ordinary Vitest runs. Per-phase
setup would let later iterations reuse an already hydrated document. The
hydration unit tests cover repeated fresh-state setup.
Each scenario is included in its framework's aggregate build and CodSpeed run. Land the benchmark independently to establish main's baseline before comparing a hydration optimization. Regenerate each revision's artifacts with identical fixture data and dependencies; do not pin an old hydration wire format forever.
Isolated route-tree construction
scenarios/route-tree-scale measures navigation over an existing tree, not
construction. scenarios/mount includes construction but also rendering and
loading. To isolate initialization, use the core construction benchmark:
CI=1 NX_DAEMON=false pnpm nx run @tanstack/router-core:test:unit --outputStyle=stream --skipRemoteCache -- bench tests/route-tree-construction.bench.ts --run --testNamePattern=1000.routes --reporter=json --outputFile=/tmp/route-tree-construction.json
It covers 100, 1,000, and 10,000 static, dynamic, nested, or mixed routes.
The processing-only case creates fresh route objects in an untimed
beforeEach, then measures processRouteTree and route.init. A separate
case includes route-object creation. Neither case measures first interpolation
or navigation. Remove the name filter to run all sizes. Compare identical
benchmark sources in separate checkouts and repeat fresh processes; a cached
router or repeatedly initialized tree is not a construction baseline.
Opt-in React Link performance suite
link-performance/ contains additional client-navigation and SSR workloads for
focused Link work. They are not included in the regular client-nav/SSR
aggregate projects or their CodSpeed build dependencies; they only run through
the dedicated @benchmarks/react-link-performance targets below. Use -t to
narrow a run to specific cases.
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/react-link-performance:test:perf:client --outputStyle=stream --skipRemoteCache -- --run
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/react-link-performance:test:perf:ssr --outputStyle=stream --skipRemoteCache -- --run
# Select a feature and save the normal Vitest JSON report.
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/react-link-performance:test:perf:client --outputStyle=stream --skipRemoteCache -- --run -t "updater|optional|splat" --reporter=json --outputFile=/tmp/link-perf.json
The cases cover repeated versus unique destination params, updater functions, relative/inherited values, search middleware chains, param stringification, optional and splat segments, encoding, masks, basepath/rewrites, and active props with structured search. These exercise different costs: cache hits and misses, parameter cloning, callbacks, middleware traversal, URI encoding, building masked/public locations, active-state comparisons, and prop merging. Existing preload, mount, and route-tree-scale scenarios remain responsible for those separate workloads.
- Client: 200 persistent measured Links, four control Links, and eight completed navigations per timed batch. The existing client harness checks hrefs and active state during its untimed warm-up lap. Control navigations replace the history entry, keeping history size constant. Post-measurement assertions also check that the measured anchors stayed mounted.
- SSR: four fresh-router requests per timed batch, each rendering 200
measured Links through
RouterProviderandrenderToString. Router creation,router.load(), rendering, and history cleanup are included. This isolates Router SSR Link work, not Start HTTP handling, dehydration, or streaming. HTML assertions run outside the timed batch. - Both use the same code-based workload definitions and production JSX/library
builds. The regular Vitest entry points use at least 100 warm-up iterations,
one second of warm-up time, and five-second measurement windows. The client
and server bundles assert their resolved
isServerenvironment. React and React DOM remain external so comparisons can share the same renderer runtime. These bundles are Node-hosted (jsdom for client mode), not browser deployments.
Run the gate tests and typecheck separately:
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/react-link-performance:test:unit --outputStyle=stream --skipRemoteCache
CI=1 NX_DAEMON=false pnpm nx run @benchmarks/react-link-performance:test:types --outputStyle=stream --skipRemoteCache
For before/after comparisons, use identical benchmark files and dependencies
on both refs, build each ref through its Nx targets, and alternate fresh
Vitest processes. Report the actual refs, per-case means and relative margins
of error; rerun noisy or borderline results with -t rather than interpreting
a small difference as a proven speedup. Client and SSR times have different
batch units and should not be compared directly.
Stable paired comparisons
For regression decisions, prefer the paired runner over a whole-file Vitest
run. It starts a fresh process for every case and repetition, avoiding JIT
feedback from earlier cases. Inside each process, both revisions share the
same production React installation but have separate router/app modules and
router instances. Initialization order alternates between repetitions.
Built snapshots are staged under node_modules/.cache so the TypeScript
harness loader does not transpile their emitted JavaScript a second time.
# Build the baseline using these same benchmark sources in its own checkout.
# --baseline points to that checkout's link-performance/dist directory.
pnpm nx run @benchmarks/react-link-performance:test:perf:stable -- \
--baseline /path/to/baseline/benchmarks/client-nav/link-performance/dist \
--outputJson /tmp/paired-links.json
# Narrow a comparison, or increase independent process repetitions.
pnpm nx run @benchmarks/react-link-performance:test:perf:stable -- \
--baseline /path/to/baseline/benchmarks/client-nav/link-performance/dist \
--mode ssr -t "middleware|unique-params" --repeats 6 \
--outputJson /tmp/paired-links-ssr.json
The runner requires Node with process.threadCpuUsage (Node 24 works).
It fixes V8 random/hash seeds, warms each variant for at least two seconds
and 100 batches, then alternates ABBA/BAAB blocks calibrated to roughly 500 ms.
Both variants do exactly the same number of batches per block. It records
main-thread CPU time, wall time, and whole-process CPU time; GC during
measurement is not disabled or discarded.
The default is four independent process repetitions. Reported 95% intervals use their paired log-ratios, not the many correlated batches as independent samples. A faster/slower verdict requires CPU and wall intervals to agree. Intervals overlapping zero or disagreeing metrics are inconclusive; narrow intervals entirely inside +/-2% are reported separately.
An A/A calibration uses the current dist directory as --baseline. Its
intervals should contain zero before trusting similarly sized A/B differences.
Shared-machine contention can still make small changes unresolved. Do not
interpret a point estimate alone, or an inconclusive result, as proof that
a workload is unchanged.