Dies ist die deutsche MFDoctor-Dokumentation. Technische Bezeichner, CLI-Flags, Regel-IDs und Codebeispiele bleiben unverändert, damit die Inhalte zwischen den Sprachen vollständig kompatibel bleiben. Verwenden Sie den Sprachumschalter für die kanonische englische Fassung.

#Performance checks

Performance advice must preserve runtime correctness. Module Federation owns part of the chunk and initialization graph, so generic bundler advice can be harmful.

#Analysebudgets

MFDoctor bounds source and workspace collection before parsing. Configure the typed analysisBudgets option when a repository needs tighter limits:

export default {
  analysisBudgets: {
    maxFiles: 10_000,
    maxSourceBytes: 50 * 1024 * 1024,
    maxArtifacts: 10_000,
    maxEvidenceNodes: 100_000,
    maxSerializedBytes: 50 * 1024 * 1024,
    maxWallTimeMs: 30_000,
  },
};

Files and bytes are selected in sorted path order. A limit stops further collection without silently claiming completeness: project analysis emits the existing doctor/partial-analysis finding and workspace analysis returns exit code 2 with unknown input. Artifact records are also selected in sorted path order and capped before JSON parsing; hitting maxArtifacts produces the same partial analysis signal. Imported evidence uses the same tracker: the node and serialized-byte reservation happens before copying, graph normalization, or stable-ID hashing. Bytes mean the UTF-8 size of the raw JSON representation, including keys and separators, and a rejected reservation is atomic. The reader throws a typed EvidenceReaderError with failureCode: "budget-exceeded" and its budget report; it does not return a clipped graph. Budget-clipped rule processing returns unknown with partial completeness, so no finding is treated as conclusive from clipped input. Legacy project/report schemas and normal runs stay unchanged. Runtime collection has a separate budget later. Projection helpers accept an optional analysisBudget and reserve the normalized input and complete output as separate atomic units; they never return silently truncated v1 data.

Source and discovered-artifact reads use a fixed worker bound and reduce results in sorted path order, so concurrency does not change file selection or report ordering. Applications that run multiple analyses in one process can opt into a bounded parsed-input cache:

import { AnalysisContentCache } from "@tonoizer/mfdoctor";

const analysisCache = new AnalysisContentCache({
  maxEntries: 256,
  maxBytes: 16 * 1024 * 1024,
});

await analyze({ analysisCache });

The cache is process-local and opt-in. Entries are invalidated by source or artifact content digest and include adapter/config identity in their key; the LRU entry and byte ceilings prevent unbounded retention. It does not replace a remote or daemon cache. The checked-in vp run benchmark:analysis command runs the small/medium/large fixtures twice through the legacy, shadow, and explicitly gate-promoted v2-compat controller selections. It records wall time/RSS/budget usage/cache reuse, checks stable v1 output, and uses the existing readEvidenceFile seam when fixture evidence is present. The v2-compat row measures the current v1 collector at the rollout-controller seam; it does not claim a separate v2 collector is enabled. Mixed Vite/Rspack/Rsbuild/Webpack coverage is provided by the existing compatibility matrix rather than this source-only benchmark.

#Startstrategie

Issue: version-first loads every configured remote entry during initialization so all shared versions are known. With many remotes, this adds startup requests and makes an offline remote an early failure.

Fix: use loaded-first when reuse of already loaded packages and on-demand remote loading matter more than choosing the highest available version. If version-first is required, add an errorLoadRemote runtime plugin with an explicit recovery policy.

Source: shareStrategy.

#Asset-Budgets

MFDoctor can fail (or warn on) federation assets whose on-disk sizes exceed project budgets. Sizes come from joining mf-manifest.json asset names to files under the project root (and emitted assets from a MFDoctor adapter build). Manifest and stats JSON do not carry byte sizes by themselves.

Default limits for performance/asset-budget:

• remote entry file: 512 KiB (524288)
• each shared package asset sum: 512 KiB (524288)
• each expose asset sum: 350 KiB (358400)

Default severity is warning. Override thresholds or severity in mfdoctor.config:

export default {
  rules: {
    "performance/asset-budget": [
      "error",
      {
        remoteEntryMaxBytes: 150_000,
        sharedMaxBytes: 400_000,
        exposeMaxBytes: 250_000,
      },
    ],
  },
};

Set the rule to "off" to disable it. When no listed asset can be sized (for example before a build), the rule reports nothing.

#Externe Laufzeit

Issue: each remote can bundle its own @module-federation/runtime-core.

Fix: a pure top consumer may enable provideExternalRuntime, while browser remotes enable externalRuntime. The provider must run first. Do not add provideExternalRuntime to a producer; the upstream plugin rejects that combination.

This can reduce duplicate runtime code, but it changes deployment order into a hard contract. MFDoctor checks both local invalid combinations and the federation-wide provider.

Source: experiments.

#Tree Shaking von Shared-Modulen

Issue: large shared component libraries can transfer exports no application uses.

Fix:

• use runtime-infer for local development and safe fallback;
• use server-calc only when a deployment service collects every consumer's used exports, builds the union, publishes the secondary artifact, and updates the snapshot;
• do not combine eager: true with shared tree shaking;
• set injectTreeShakingUsedExports: false for server-calc.

If any deployment step is missing, the runtime falls back to the full package. That is safe but means the expected size win did not happen.

Sources: shared, treeShakingDir, and the Vite implementation notes.

#Entfernen von Laufzeitfähigkeiten

Vite can remove unused remote, shared, and snapshot runtime features. Core exposes matching optimization flags under experiments.optimization.

• disableRemote is only safe with no remotes.
• disableShared is only safe with no shared packages.
• disableSnapshot removes manifest remotes, preload, dynamic type hints, HMR, and DevTools integration.

MFDoctor reports a hard error when a removed capability is still configured. It reports snapshot loss as reliability risk because the build may run while important tooling silently disappears.

#Vite-specific costs

• bundleAllCSS: true attaches all CSS to every expose. Use it only when each expose needs the full style set.
• Large projects should prefer moduleParseIdleTimeout over a short fixed moduleParseTimeout.
• User manualChunks and custom codeSplitting.groups are ignored by the official plugin because they can break federation initialization order. Let the plugin isolate loadShare and runtime-init chunks.

Source: Vite plugin source.