Files
clean-architecture-frontend…/vite.config.ts
T
DongHyeonka 3f9c087798 docs(performance): 초기 청크 수의 HTTP/1.1 비용을 재고 묶지 않는 근거를 적는다 (1101/14.4)
문제
08-08은 초기 JS 바이트를 줄이는 대신 초기 파일을 16개에서 25개(entry
포함 26개)로 늘렸고, "HTTP/1.1 호스팅이면 다시 본다"를 남겼다. vite
preview도 HTTP/1.1이다. 비용을 잰 적이 없어 위험이 열려 있었다.

고른 방법
- 지금 빌드와 묶은 빌드 두 가지를 쟀다. (가) 08-08이 나눈 초기 primitive
  청크 4개(field·button·select·page-templates)를 하나로. (나) 묶을 수 있는
  초기 청크 14개를 하나로(entry·preload-helper·boot, 부트 실패 화면 closure,
  15-01 그룹은 뺐다). 같은 커밋을 임시 복사본에서 그룹 하나만 더해 빌드했다.
- 측정: vite preview(HTTP/1.1, gzip), Chromium CDP 네트워크 에뮬레이션
  (DevTools Fast 3G·Fast 4G 값), 새 context·캐시 없음, 첫 화면 =
  boot 셸이 사라지고 route h1이 들어온 시각. 세 빌드를 번갈아 7번씩, 중앙값.
  HTTP/2(node:http2, 자체 서명 인증서, gzip)도 같은 방법으로 쟀다.
- 결과(Fast 3G / Fast 4G 중앙값)
  - 지금 26개 202,782 B: 5,887 / 1,874 ms
  - (가) 23개 201,818 B: 5,426 / 1,764 ms
  - (나) 13개 198,673 B: 4,930 / 1,535 ms
  - HTTP/2: 지금 4,793 / 1,444, (가) 4,769 / 1,440, (나) 4,708 / 1,422 ms
- 판단: 묶지 않는다. 측정값과 판단을 performance-evidence.md에 적고,
  vite.config.ts의 "재검토" 주석이 그 절을 가리키게 한다.
  - (가)·(나)는 이 빌드의 청크 안 모듈 목록을 그대로 적어 만들었다. 규칙으로
    고를 방법이 없었다. boot 정적 closure는 design-system 배럴을 지나 갤러리
    전용 primitive까지 잡는다. rolldown `$initial` 태그는 entry가 앱을 dynamic
    import로 읽어서(09-03) 아무것도 잡지 않는다. boot.tsx를 entry로 더하면
    manifest 분류가 lazy 화면까지 초기로 센다(28개, 216,820 B, 예산 초과).
  - 고정 목록은 셸이 쓰는 primitive가 바뀔 때마다 맞춰야 하고, 낡아도
    실패하는 검사가 없다. 08-08이 같은 이유로 manualChunks 이름 고정을 버렸다.
  - 템플릿은 HTTP/2 이상 서빙을 전제로 본다. HTTP/1.1로만 서빙하는 제품은
    (나) 같은 그룹을 넣고 셸을 바꿀 때마다 다시 잰다고 문서에 적었다.
버린 대안
- (나)를 고정 목록으로 넣기: HTTP/1.1에서 Fast 3G 0.96초 빠르지만 숨은
  결합이다(위).
- (가)만 넣기: 0.46초. 같은 이유.

위험·한계
- HTTP/1.1로 서빙하는 다운스트림은 첫 화면이 Fast 3G에서 약 1초 늦다.
  이 커밋은 그 비용을 없애지 않고 숫자와 전제를 적는다. 호스팅 HTTP 버전을
  검사하는 게이트는 없다(FE-GATE-019 probe 밖).
- 측정은 이 머신의 loopback과 CDP 에뮬레이션이다. 연결 수립(TCP·TLS) 비용은
  CDP 지연에 들어가지 않는다. loadavg는 HTTP/1.1 측정 6~20, HTTP/2 측정
  3~6이라 두 표끼리는 비교하지 않는다. 각 표 안은 번갈아 재서 비교할 수 있다.

검증
- 빌드 바이트: classify-vite-bundle + gzipSync(check:bundle과 같은 계산).
  지금 빌드는 기준 202,770 B(local-build)와 맞다.
- 측정 스크립트와 로그는 커밋하지 않았다(worktree의 .tmp/ra/, 일회성).
- node scripts/verify-documentation-readiness.ts: PASS_SCOPED.
- npx eslint vite.config.ts, npx tsc --project tsconfig.node.json --noEmit 통과.
2026-11-03 20:47:52 +09:00

210 lines
8.6 KiB
TypeScript

import path from "node:path";
import { defineConfig, loadEnv } from "vite";
import react from "@vitejs/plugin-react";
import tailwindcss from "@tailwindcss/vite";
import { viteModuleInventoryPlugin } from "./scripts/lib/vite-module-inventory.ts";
import { ENV_REGISTRY } from "./src/contracts/env.ts";
/**
* 13-06. Every VITE_* value Vite loads (shell or `.env*`) is public once
* something bundles it, so a value the build does not declare is refused
* rather than carried along.
*/
function assertDeclaredBuildVariables(mode: string): void {
const undeclared = Object.keys(loadEnv(mode, process.cwd(), "VITE_"))
.filter((name) => !Object.hasOwn(ENV_REGISTRY, name))
.sort();
if (undeclared.length > 0) {
throw new Error(
`Undeclared VITE_* build variables: ${undeclared.join(", ")}. ` +
"Declare them in ENV_REGISTRY (src/contracts/env.ts) or unset them; " +
"VITE_ values are published to the browser.",
);
}
}
/**
* §6.1. One sub-path, declared once.
*
* `VITE_ROUTER_BASE_PATH` already drives the router and the Service Worker
* scope. Vite's asset `base` was left at its default, so a build served from
* `/app/` emitted root-absolute asset URLs and loaded nothing: the three
* consumers of the same setting disagreed. They are read from one value here so
* a sub-path deployment is coherent or fails at build time.
*/
function routerBasePath(environment: NodeJS.ProcessEnv): string {
const declared = environment["VITE_ROUTER_BASE_PATH"];
if (declared === undefined || declared === "") return "/";
if (!declared.startsWith("/") || !declared.endsWith("/")) {
throw new Error(
`VITE_ROUTER_BASE_PATH must start and end with "/"; received ${declared}`,
);
}
return declared;
}
/**
* 13-07. The runtime document ships inside dist, which is served under the
* base path. An explicit root-relative URL outside it would 404 at boot.
*/
function assertRuntimeConfigUnderBase(
environment: NodeJS.ProcessEnv,
base: string,
): void {
const declared = environment["VITE_RUNTIME_CONFIG_URL"];
if (declared?.startsWith("/") && !declared.startsWith(base)) {
throw new Error(
`VITE_RUNTIME_CONFIG_URL ${declared} is outside VITE_ROUTER_BASE_PATH ${base}`,
);
}
}
type ModuleGraph = Readonly<{
getModuleInfo(id: string): Readonly<{
importedIds: readonly string[];
dynamicallyImportedIds: readonly string[];
}> | null;
}>;
function staticClosure(
graph: ModuleGraph,
roots: readonly string[],
within: (id: string) => boolean = () => true,
): Set<string> {
const reached = new Set<string>();
const pending = [...roots];
for (let next = pending.pop(); next !== undefined; next = pending.pop()) {
if (reached.has(next) || !within(next)) continue;
reached.add(next);
pending.push(...(graph.getModuleInfo(next)?.importedIds ?? []));
}
return reached;
}
/**
* 15-01. Two chunk groups that keep moving the feature runtime halves out of
* the initial JavaScript from costing more than it saves.
*
* - `feature-declarations`: every installed feature's declaration, the static
* graph under src/features/ that the generated feature index reaches
* (feature-module.ts, its route contract and messages). A declaration is
* shared by the boot graph and its own feature's lazy runtime and pages,
* and rolldown otherwise gives each such set of importers a chunk of its
* own: one more initial file per feature, which doubled what a feature
* added to the initial JavaScript.
* - `feature-platform`: the boot modules the runtime halves also import (HTTP
* executor, contract runtime, zod). Left to rolldown they split into one
* initial chunk per set of importers; the per-file gzip overhead and the
* compression lost across files cost more than the runtime code removed.
* Modules that the boot-failure screen needs are left out, so
* render-boot-failure.tsx never depends on this chunk: it must still render
* when a module of the application graph throws while evaluating (09-03).
*
* Both only regroup modules the boot already loads; nothing lazy becomes
* initial. If feature-declarations comes out empty, the per-feature bound in
* tests/unit/feature-shape-growth.test.ts fails. feature-platform only moves
* the platform total (about 2 KB at 15-01), which the bundle budget
* (FE-GATE-012) checks.
*/
function createFeatureChunkGroups(root: string) {
const at = (file: string) => path.join(root, file);
const featuresRoot = `${at("src/features")}${path.sep}`;
let groups: ReadonlyMap<string, string> | undefined;
const classify = (graph: ModuleGraph): ReadonlyMap<string, string> => {
const index = graph.getModuleInfo(at("src/.generated/product-feature-index.ts"));
const declarations = staticClosure(graph, index?.importedIds ?? [], (id) =>
id.startsWith(featuresRoot),
);
const runtimes = staticClosure(
graph,
[...declarations].flatMap((id) => graph.getModuleInfo(id)?.dynamicallyImportedIds ?? []),
);
const boot = staticClosure(graph, [at("src/bootstrap/boot.tsx")]);
const bootFailure = staticClosure(graph, [at("src/bootstrap/render-boot-failure.tsx")]);
const result = new Map<string, string>();
for (const id of declarations) result.set(id, "feature-declarations");
for (const id of runtimes) {
if (boot.has(id) && !bootFailure.has(id) && !declarations.has(id)) {
result.set(id, "feature-platform");
}
}
return result;
};
return (id: string, graph: ModuleGraph): string | null => {
groups ??= classify(graph);
return groups.get(id) ?? null;
};
}
export default defineConfig(({ mode }) => {
assertDeclaredBuildVariables(mode);
const base = routerBasePath(process.env);
assertRuntimeConfigUnderBase(process.env, base);
return {
base,
plugins: [react(), tailwindcss(), viteModuleInventoryPlugin()],
build: {
manifest: true,
sourcemap: false,
rollupOptions: {
/**
* 1101/08-08. The design-system modules have no top-level effects:
* they declare components, icons and token lists. Declared so, a
* primitive file none of whose exports a chunk uses is dropped from
* it, instead of riding the initial chunk because the app shell
* imports the index that re-exports it. The bundler drops an import
* kept only for its effect along with its module, so these files hold
* no such import; tests/unit/design-system-initial-chunk.test.ts
* checks both. Stylesheets load from bootstrap/main.tsx.
*
* Trade: at 08-08 the initial JavaScript fell by 2,718 gzip bytes
* and the initial chunks rose from 16 to 25. The entry preloads them
* in parallel, so over HTTP/2 they add no round trip. Over HTTP/1.1
* they queue for six connections; 1101/14.4 measured that cost and
* why the chunks stay split ("초기 청크 수와 HTTP/1.1" in
* docs/operations/performance-evidence.md).
*/
treeshake: {
moduleSideEffects: [
{
test: /[\\/]src[\\/]presentation[\\/]design-system[\\/].+\.tsx?$/,
sideEffects: false,
},
],
},
output: {
codeSplitting: {
groups: [
// 09-03. The entry now imports the app dynamically, so Vite's
// preload helper would land in the entry chunk and every lazy
// chunk would import the entry, whose hash changes with each
// build id.
{ name: "preload-helper", test: /vite[\\/]preload-helper/ },
// 1101/07-03. Every feature's schemas and runtime half import
// the shared zod adapter. Alone it becomes a chunk of its own
// once two features share it, and every route and runtime
// import then preloads one more file (about 50 gzip bytes of
// initial JavaScript for the second feature). Those imports
// already preload schema-registry.ts, so the adapter rides
// that lazy chunk. Both modules stay off the boot graph.
{
name: "schema-registry",
test: /[\\/]src[\\/]contracts[\\/](?:schema-registry|zod-validator)\.ts$/,
includeDependenciesRecursively: false,
},
{
name: createFeatureChunkGroups(process.cwd()),
// The boot-failure screen's modules stay out of
// feature-platform even when a grouped module imports them.
includeDependenciesRecursively: false,
},
],
},
},
},
},
};
});