refactor: replace cachetools with stdlib OrderedDict-based LRUCache

[isaacbmiller]

Summary

Remove cachetools as a direct dependency by replacing cachetools.LRUCache with a ~30-line OrderedDict subclass.

Changes

• dspy/clients/cache.py: Added LRUCache(OrderedDict) class covering maxsize, get/set/contains/clear, LRU eviction via move_to_end, and cloudpickle serialization via __reduce__.
• tests/clients/test_cache.py: Updated import from cachetools to dspy.clients.cache.
• pyproject.toml: Removed cachetools>=5.5.0 from dependencies.

Notes

• Thread safety is handled by the existing self._lock = threading.RLock() in the Cache class, not in LRUCache itself.
• All 21 cache tests pass, including save/load via cloudpickle.

[greptile-apps]

Greptile Summary

This PR removes the cachetools direct dependency by replacing cachetools.LRUCache with a ~30-line OrderedDict subclass that covers the exact API surface used by DSPy.

• dspy/clients/cache.py: New LRUCache(OrderedDict) correctly implements LRU promotion via move_to_end in both __getitem__ and __setitem__, evicts least-recently-used items when maxsize is exceeded, and supports cloudpickle round-trips via a __reduce__ 5-tuple that re-inserts items into a fresh instance.
• pyproject.toml / uv.lock: cachetools>=5.5.0 removed from direct dependencies and lock file; thread safety is unchanged, delegated to the existing RLock in Cache.
• tests/clients/test_cache.py: Import updated to use the new in-package LRUCache; all 21 existing tests continue to exercise the same behaviors.

Confidence Score: 5/5

Safe to merge — the custom LRUCache correctly replicates the cachetools contract used by DSPy, cloudpickle serialization is preserved via reduce, and thread safety is unaffected.

The replacement implementation faithfully covers every code path the Cache class exercises: LRU promotion on read and write, front-of-dict eviction when capacity is exceeded, and pickle round-tripping. The existing 21-test suite covers memory cache get/put, save/load, and eviction. No behavioral regression is apparent.

No files require special attention.

Important Files Changed

Filename	Overview
dspy/clients/cache.py	Adds a custom LRUCache(OrderedDict) subclass replacing cachetools.LRUCache; implementation correctly handles get/set with LRU promotion, eviction, and cloudpickle serialization via reduce.
tests/clients/test_cache.py	Updated import from cachetools.LRUCache to dspy.clients.cache.LRUCache; test coverage unchanged and all existing cases still apply to the new implementation.
pyproject.toml	Removes cachetools>=5.5.0 from direct dependencies; the dependency is no longer needed.
uv.lock	Lock file updated to remove cachetools package entry; consistent with pyproject.toml change.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["LRUCache.__setitem__(key, value)"] --> B{key in cache?}
    B -- Yes --> C["move_to_end(key)\n(promote to MRU)"]
    C --> D["super().__setitem__(key, value)\n(update in-place)"]
    B -- No --> D
    D --> E{len > maxsize?}
    E -- Yes --> F["pop(next(iter(self)))\n(evict LRU item)"]
    F --> E
    E -- No --> G[Done]

    H["LRUCache.__getitem__(key)"] --> I["move_to_end(key)\n(promote to MRU)"]
    I --> J["return super().__getitem__(key)"]

    K["LRUCache.__reduce__()"] --> L["return (LRUCache, (maxsize,), None, None, iter(items))"]
    L --> M["Pickle reconstructs:\nLRUCache(maxsize)\nthen obj[k]=v for each item"]

Loading

_{Reviews (3): Last reviewed commit: "refactor: replace cachetools with stdlib..." | Re-trigger Greptile}