Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -48,37 +48,39 @@ const TokenDisplay = ({ token }: TokenDisplayProps) => {

**DO:**

- Reduce dependencies by moving values to default parameters when possible
- Only include dependencies that actually trigger the effect
- Depend on the primitive the effect reads (an ID, an address, a number) rather than an object rebuilt on each render
- Keep values the effect does not use out of the effect

**DON'T:**

- Include unnecessary dependencies that cause effects to run too often
- Drop a value the effect reads to make it run less often. A default parameter does not stop it being a dependency (see Rule: Include All Dependencies in useEffect)

**Example - WRONG:**

```typescript
const TokenBalance = ({ address, network, refreshInterval }: Props) => {
const TokenBalance = ({ account, network, refreshInterval }: Props) => {
const [balance, setBalance] = useState('0');

useEffect(() => {
const fetch = async () => {
const result = await fetchBalance(address, network);
const result = await fetchBalance(account.address, network);
setBalance(result);
};

fetch();
const interval = setInterval(fetch, refreshInterval);
return () => clearInterval(interval);
}, [address, network, refreshInterval]); // Effect runs too often
}, [account, network, refreshInterval]); // A parent passing account={{ address }} rebuilds it on every render, so the effect refetches and resets the interval each time
};
```

**Example - CORRECT:**

```typescript
const TokenBalance = ({ address, network, refreshInterval = 10000 }: Props) => {
const TokenBalance = ({ account, network, refreshInterval }: Props) => {
const [balance, setBalance] = useState('0');
const { address } = account;

useEffect(() => {
const fetch = async () => {
Expand All @@ -89,7 +91,7 @@ const TokenBalance = ({ address, network, refreshInterval = 10000 }: Props) => {
fetch();
const interval = setInterval(fetch, refreshInterval);
return () => clearInterval(interval);
}, [address, network]); // refreshInterval moved to default param
}, [address, network, refreshInterval]); // The address string is equal across renders
};
```

Expand Down Expand Up @@ -344,7 +346,7 @@ const useHistoricalPrices = () => {
**DO:**

- Split effects when conditional logic excludes some dependencies
- Ensure all dependencies in array are actually used
- Ensure the effect reads every dependency (see Rule: Don't List Dependencies the Effect Doesn't Read)

**DON'T:**

Expand Down Expand Up @@ -374,19 +376,61 @@ const useHistoricalPrices = ({ isEvm, chainId, address }: Props) => {
fetchPrices(chainId, address);
}
}, [isEvm, chainId, address]); // All deps are used
};
```

### Rule: Don't List Dependencies the Effect Doesn't Read

The dependency array describes the effect's code. A value that must restart the effect when it changes has to be a value the effect uses.

**DO:**

// OR Option 2: Separate effects
- Treat a dependency the effect never reads as a finding, even when the effect is meant to restart when it changes
- If the effect's work depends on the value, pass the value into that work: the call that starts the connection, subscription or background registration
- If the value marks a scope whose state must all reset when it changes, render that scope as a component and give it the value as its `key`
- If the value changes because of a specific action, do the work where that action is handled instead of in an effect that watches for the change

**DON'T:**

- Delete the dependency without one of the changes above. The effect then keeps a session bound to the old value
- Keep the unread dependency and explain it in a comment. React Compiler's effect-dependency validation (`validateExhaustiveEffectDependencies`, lint rule `react-hooks/exhaustive-effect-dependencies`, off by default) reports it as unnecessary, and with that validation on, the compiler leaves the whole component or hook uncompiled, or fails the build under a strict `panicThreshold`
- Add `'use no memo'` to protect the dependency array. It opts the whole function out of React Compiler, and React documents it as a temporary debugging tool
- Add a parameter the called code ignores just so the effect reads the value. That moves the unread value into the call

**Reference:** [Removing Effect Dependencies](https://react.dev/learn/removing-effect-dependencies), [Resetting all state when a prop changes](https://react.dev/learn/you-might-not-need-an-effect#resetting-all-state-when-a-prop-changes), [`"use no memo"`](https://react.dev/reference/react-compiler/directives/use-no-memo)

**Example - WRONG:**

```typescript
const useSession = ({ address, networkId }: Props) => {
useEffect(() => {
if (isEvm) return;
fetchPrices(chainId, address);
}, [isEvm]); // Only depends on condition
const session = startSession(address);
return () => session.stop();
}, [address, networkId]); // networkId is never read
};
```

**Example - CORRECT:**

```typescript
// The session is for one network, so the effect passes the network in
const useSession = ({ address, networkId }: Props) => {
useEffect(() => {
if (!isEvm) {
fetchPrices(chainId, address);
}
}, [chainId, address]); // Only when not EVM
const session = startSession({ address, networkId });
return () => session.stop();
}, [address, networkId]);
};

// Caches must be cleared when the account or network changes:
// a keyed child unmounts on the change, and its cleanup clears them
const CacheScope = () => {
useEffect(() => () => clearCaches(), []);
return null;
};

const Root = ({ address, networkId }: Props) => (
<CacheScope key={`${address}:${networkId}`} />
);
```

### Rule: Use useRef for Persistent Values
Expand Down Expand Up @@ -856,6 +900,76 @@ const PriceTicker = ({ tokenAddress }: PriceTickerProps) => {
};
```

### Rule: Keep Multi-Step Async Protocols Out of the Effect

**DO:**

- When an effect acquires an external resource through more than one async step, with a timeout or more than one way to end, move that protocol into a plain object owned outside React: `start()` returns a handle, and `handle.stop()` releases the resource
- Keep the effect to acquiring and releasing: start in the body, stop in the cleanup
- Unit-test the object's end paths directly (ready, timeout, failure, stop while a step is in flight), without rendering

**DON'T:**

- Grow a state machine of closure flags (`cancelled`, `ended`, `ready`) inside the effect. Every end path needs its own cleanup, every fix has to reason about every flag at once, and the only way to test it is through `renderHook`

**Why:** every line of such an effect can pass the rules above (stable dependencies, timers cleared, a cancelled flag on the async chain) while the bugs sit between its end paths: a timeout that fires mid-acquire, a release sent before registration, a release that also runs on the success path. Those bugs belong to the protocol, not to any one line, so a line-level review does not catch them.

**Example - WRONG:**

```typescript
useEffect(() => {
let cancelled = false;
let ended = false;
let ready = false;
const finish = (reason: string) => {
if (!ended) {
ended = true;
report(reason);
}
};
const timeout = setTimeout(() => {
cancelled = true;
finish('timeout');
release(id);
}, 30_000);
register(id)
.then(() => !cancelled && init(address))
.then(() => !cancelled && start(id))
.then(() => {
if (!cancelled) {
ready = true;
finish('ready');
}
})
.catch(() => {
finish('failed');
if (!cancelled) release(id);
})
.finally(() => clearTimeout(timeout));
return () => {
cancelled = true;
clearTimeout(timeout);
finish('released');
release(id, ready);
};
}, [address]);
```

**Example - CORRECT:**

```typescript
// preload-session.ts: a plain object, unit-tested without React
export function startPreloadSession(address: string): { stop: () => void } {
// register, init, start, the timeout, and every end path live here
}

// The hook only acquires and releases
useEffect(() => {
const session = startPreloadSession(address);
return () => session.stop();
}, [address]);
```

### Rule: Avoid Large Object Retention in Closures

**DO:**
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -694,6 +694,46 @@ const TokenList = ({ tokens, filter }: TokenListProps) => {
};
```

### ❌ Anti-Pattern: Dependencies the Effect Never Reads

**Problem:** An entry the effect does not read is not a harmless extra. The compiler's
effect-dependency validation treats it as an error and **skips the whole function**, so the
hook ships with no memoization at all. A comment next to the array does not help: the check
reads the array, not the prose.

**Solution:** Make the value one the effect actually reads, by passing it into the work the
effect starts. Where it marks a scope that must reset as a whole, put a `key` on a component
instead. Never satisfy it with a comment, with `'use no memo'`, or with a parameter the callee
ignores.

**Example - WRONG:**

```typescript
useEffect(() => {
const session = manager.startPreload({ address, accountChanged });
return () => session.stop();
}, [
address,
// Provider or network changes release the old preload and register a new one.
provider,
isTestnet,
]);
```

**Example - CORRECT:**

```typescript
useEffect(() => {
// `provider` and `isTestnet` are now read, so the array describes the code.
const session = manager.startPreload({ address, accountChanged, provider, isTestnet });
return () => session.stop();
}, [address, provider, isTestnet]);
```

**Why this lives here as well as in `perf-hooks-effects`:** the rule is stated there as an
effects rule, and the cost is a React Compiler cost. Someone asking what loses them compiler
coverage opens this file, so the rule has to be reachable from this side too.

### ❌ Anti-Pattern: Using Index as Key for Dynamic Lists

**Problem:** Breaks React's reconciliation when lists can be reordered, filtered, or have items added/removed.
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
---
title: React Compiler Error Triage & Coverage Accounting (MetaMask)
impact: HIGH
tags: react-compiler, panicThreshold, error-triage, coverage, babel, build
---

# Skill: React Compiler Error Triage & Coverage Accounting

The React Compiler **fails open**: when it can't compile a component, it silently skips it and ships the unoptimized original. The build stays green, DevTools shows no warning — you just don't get the memoization. Once the compiler is enabled broadly (metamask-mobile#31171 enabled v1.0.0 app-wide), the question stops being "is it on?" and becomes **"what is it actually compiling, and which of its errors are worth fixing?"** This file is the triage playbook. Extension PR metamask-extension#38007 is the reference implementation.

## The `panicThreshold` ladder

`panicThreshold` controls when a compiler diagnostic fails the build instead of silently skipping the file:

| Setting | Build fails on | Use for |
|---|---|---|
| `'none'` (default) | never — every failed file is **silently skipped** | production builds, always |
| `'critical_errors'` | only critical errors (compiler-internal invariant violations) | CI / debug builds, first ratchet target |
| `'all_errors'` | every diagnostic, including unsupported syntax | CI / debug builds, end-state ratchet |

**The ratchet strategy** (extension roadmap): keep production at `'none'` permanently; aim for a *non-production* build that passes at `'critical_errors'`, then at `'all_errors'`. Each ratchet step turns a class of silent skips into a visible, fixable error list. Never enable a non-`'none'` threshold in a release build — one un-compilable file would block the release for an optimization that is optional by design.

## Triage: unsupported syntax vs. legitimate errors

Compiler diagnostics are **not one bucket**. The logger event's `category` field separates them, and the distinction decides whether you act:

- **`category === 'Todo'` → "unsupported."** Syntax or a pattern the compiler *itself* has not implemented yet. There is **no actionable fix on our side** — rewriting working code to appease an unimplemented compiler path is wasted effort and churn. Count these separately, leave the code alone, and re-check after compiler upgrades.
- **Any other category (e.g. `InvalidReact`, `InvalidJS`) → legitimate, actionable.** A real Rules-of-React violation in our code (mutation during render, conditional hooks, side effects in render). Fixing it both unlocks compilation *and* removes a latent correctness bug.

A healthcheck that doesn't make this split is noise: the `Todo` count swamps the actionable list and the team learns to ignore the output. The extension's verbose run at enablement (metamask-extension#38007) is the canonical illustration — of 7,308 files processed: 253 compiled, **31 actionable errors**, **7,024 unsupported** (`Todo`). Without the split that reads as ~7,000 hopeless errors; with it, the team's backlog is 31 files and the rest is the compiler's to burn down across upgrades. The extension's webpack wrapper makes the split in ~10 lines:

```ts
// adapted from metamask-extension development/webpack/utils/loaders/reactCompilerLoaderWrapper.ts
// (mobile equivalent: pass a `logger` in babel-plugin-react-compiler options)
logger: {
logEvent(filename, event) {
switch (event.kind) {
case 'CompileSuccess': record(filename, 'compiled'); break;
case 'CompileSkip': record(filename, 'skipped'); break;
case 'CompileError': {
const category = event.detail?.options?.category ?? event.detail?.category;
// 'Todo' = not yet supported by the compiler — no actionable fix on our side
record(filename, category === 'Todo' ? 'unsupported' : 'error');
break;
}
}
},
}
```

The extension exposes this as `yarn webpack --reactCompilerVerbose` (per-file ✅/⏭️/🔍/❌ output + summary stats) and `--reactCompilerDebug={all|critical|none}` (maps to `panicThreshold: '<value>_errors'`). On mobile the same taxonomy is available through the Babel plugin's `logger` option or `eslint-plugin-react-compiler` (the lint rule runs the same analysis the compiler does).

## Coverage accounting

Track four buckets — **compiled / skipped / errors / unsupported** — at file and component granularity, with **worst-status-wins per file** (`error > unsupported > skipped > compiled`): a file with five compiled components and one error is an *error file*, otherwise mixed files inflate the compiled count and the number lies to you.

What the buckets tell you:

- **compiled** — your real optimization coverage, minus one gap: a module-scope `'use no memo'` directive still logs `CompileSuccess` for every function in the file before the directive discards the transform, so those files land here too, not in `skipped`. "The compiler is enabled" claims nothing; this number does.
- **errors** — the actionable backlog. Each is a Rules-of-React fix.
- **unsupported** — the compiler's backlog, not yours. Trend it across compiler upgrades.
- **skipped** — intentional exclusions: test/story files and **class components** (never compiled — metamask-mobile#30919 counted 53 at full enablement; migration to function components is the only way to move them into the compiled bucket). A `'use no memo'` directive lands here only when it sits inside one function's body, logging `CompileSkip`. A module-scope one, at the top of the file, logs as `compiled` instead.

## Staged adoption roadmap

The extension's sequence generalizes to any repo:

1. **Lint clean:** update `eslint-plugin-react-hooks` / `eslint-plugin-react-compiler` to latest; fix violations — these are exactly what the compiler will refuse to compile.
2. **Audit opt-outs:** every `'use no memo'` carries a reason + TODO; the count only goes down. `grep -rn "use no memo" app --include="*.ts*"`. A directive can be masking nothing: remove it and re-run the compiler, and zero new errors means it was not load-bearing.
3. **Ratchet `critical_errors`:** non-prod build passes; fix what surfaces.
4. **Ratchet `all_errors`:** remaining actionable errors fixed; what's left is the `Todo` (unsupported) set, which you wait out.

## Verify

- Per component: `Memo ✨` badge in React DevTools (see [js-profile-react.md](js-profile-react.md)).
- Per repo: the compiled-files count from the logger stats rises (or at least doesn't silently fall) release over release, read against the `'use no memo'` opt-out count from the roadmap's audit step. A module-scope directive still logs as `compiled`, so the raw count alone can rise while real coverage doesn't. Silent coverage regressions are the failure mode this file exists to catch.
- After a compiler version bump: re-run the verbose build and diff the `unsupported` list — `Todo`s that became `compiled` are free wins; new `error`s are regressions to triage.

## Don't over-correct

- **Never "fix" a `Todo`.** Rewriting working code around an unimplemented compiler feature is churn with no perf evidence; the next compiler release may compile it as-is.
- Don't gate releases on compiler errors (`panicThreshold` stays `'none'` in production builds).
- Don't treat `skipped` as a problem — tests, stories, and function-body opt-outs belong there (a module-scope `'use no memo'` counts as `compiled`, not `skipped`). The smell is *unexplained* `'use no memo'` directives, not the bucket itself.
- A component without `Memo ✨` is not automatically a bug to chase — check the buckets first; it may be `unsupported`.

## Related

- [mm-react-compiler.md](mm-react-compiler.md) — enabling the compiler in this repo (Babel config, Metro cache, ESLint healthcheck)
- [js-react-compiler.md](js-react-compiler.md) — how the compiler transforms code; Rules-of-React background
- [mm-selector-cascade.md](mm-selector-cascade.md) — what the compiler **cannot** fix: unstable values crossing file boundaries (selectors, imported hooks)
Loading
Loading