Skip to content

Cache keys are hashed absolute paths, so the cache can never be reused across checkouts (CI → local, git worktrees) #8488

Description

@webard
Subject Details
Rector version 2.6.6
Installed as composer dependency

ChangedFilesDetector keys every cache entry on the absolute path of the file:

private function getFilePathCacheKey(string $filePath): string
{
    return $this->fileHasher->hash($this->resolvePath($filePath) . $this->scopeSuffix);
}

private function resolvePath(string $filePath): string
{
    $realPath = realpath($filePath);
    // ...
    return $realPath;
}

The path is hashed into the cache file's name, so the cache is bound to one location on disk. Move the project - a second git worktree, a CI checkout, a container mount - and the cache is 100% misses while containing nothing that could be rewritten to fix it.

Measurements

A 12 762-file project, 14 cores, PHP 8.5.10.

Scenario Wall time
fresh checkout, empty cache 224 s
same checkout, warm cache 5 s
fresh checkout, cache copied from another checkout of the same commit 247 s (291 MB copied for nothing)

For comparison, the same experiment with the other two tools in the same toolchain, copying their caches into the same fresh checkout:

Tool cold seeded from another checkout
Pint (PHP-CS-Fixer) 105 s 2.7 s
PHPStan 43 s 4.4 s
Rector 224 s 247 s

Rector's cache records nothing about the environment. Grepping all 12 761 entries of a populated tmp/rector finds zero references to a PHP version, an extension list or an OS - each entry is just a content hash. The absolute path is the only thing standing between that cache and reuse somewhere else.

Proposal

Key on the path relative to the project root and re-absolutize on read, as PHPStan does.
Two honest caveats:

  1. Existing caches invalidate once on upgrade.
  2. Files outside the project root (if any can reach the detector) need a defined behaviour - PHPStan's transformer leaves an already-absolute path untouched, which also keeps old-format caches readable.

I am happy to send a PR if you agree with the direction. I would rather hear which shape you want (anchor directory as a constructor dependency vs resolving it from the config) than guess and have it rewritten.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions