perf: config classes check every property against the environment on each construction #10615

Problem

BaseConfig::__construct() checks every property of a config class against the environment. It also walks into arrays, down to each leaf value. For each leaf, getEnvValue() tries:

• 4 array_key_exists() checks on $_ENV
• 4 array_key_exists() checks on $_SERVER
• then, when nothing matched, 4 getenv() calls

Most properties have no env override. So the most common path is also the slowest: 8 failed lookups plus 4 failed getenv() calls for every leaf. The cost grows with the number of config properties, not with the number of env overrides.

The .env file itself is parsed only once, by DotEnv. The waste is the per-property lookup, not the parsing.

This matters most in two cases:

• Requests without config caching (FactoriesCache).
• FrankenPHP Worker Mode. Factories::reset() runs on every request, so configs are rebuilt, and pay this cost, on every request.

Measurements

PHP 8.3 CLI on macOS. 20 core config classes (App, Database, Mimes, Cache, Logger, Security, Session, Cookie, Filters, Routing, Toolbar, Email, Exceptions, Paths, Kint, Format, Validation, View, Feature, Encryption). That's 287 leaf values. Each number is the time to build all 20, averaged over 500 runs.

Proposal

Scan the environment once and group keys by the text before their first . or _. Each config class then takes only the keys that match its short prefix (app) or full prefix (Config\App).

• If a class has no matching keys, skip the property walk entirely. This is the common case.
• If it has matching keys, run today's walk unchanged, but look up only that class's small key set. No getenv() calls.

The current matching rules stay the same:

1. The short prefix wins over the full prefix.
2. $_ENV wins over $_SERVER, which wins over getenv().
3. The dot form wins over the underscore form.
4. Arrays only accept overrides for keys that already exist in the default.

Staying fresh

The index keeps snapshots of $_ENV and $_SERVER. It compares them with === on each construct.

• PHP arrays are copy-on-write. When nothing has written to them, the snapshot and the superglobal are the same array in memory, so the check takes ~22 ns.
• Any write breaks the match, and the index rebuilds. The rebuild also re-reads getenv() (~3.3 µs for ~100 vars).
• BaseConfig::reset() clears the index.

In Worker Mode, $_SERVER is refreshed on every request. So the index rebuilds once per request, on its own. The worker template needs no change.

Compatibility

• Overridden hooks: if a subclass overrides getEnvValue() or initEnvValue(), that class falls back to today's full walk with live getenv() reads. Detection uses reflection once per class, cached for the process. Those classes behave exactly as they do today.
• New private state only. No new public or protected API.

Estimated gains

Prototype measured with the same setup as above. "Typical overrides" means app.baseURL plus four database.default.* keys.

In absolute terms, this saves about 0.4 ms per request for these 20 classes. Apps that load more config classes, or classes with large arrays, will save more.

Caveats on these numbers:

• They come from a scratch prototype, not the final code.
• The prototype skipped Registrar handling in its timings. The current-code timings include it. With no Registrars, that's a small cost, but it makes the prototype look slightly better than the real thing will.
• They come from a single dev machine. Opcache and JIT settings will shift them.
• Requests served from config caching (FactoriesCache) already skip this work. They won't gain anything.

Behavior changes (4.8)

1. A bare putenv() call, with no matching $_ENV/$_SERVER write, after the first config is built is not seen until the snapshot changes or BaseConfig::reset() runs.
2. On Windows, getenv() ignores case. Env vars that exist only in the OS environment will now match case-sensitively. Values from .env already matched case-sensitively through $_ENV/$_SERVER.

Both will be listed in the v4.8.0 changelog under "Behavior Changes", with a note in upgrade_480.rst.

Out of scope (possible follow-ups)

• registerProperties() creates a ReflectionClass on every construct just to get the short class name.
• The Encryption instanceof check runs once per property inside the constructor loop.