From 04a76b8e154f5c2be951ab38f949b8eddfcde804 Mon Sep 17 00:00:00 2001
From: emme1t <149944796+emme1t@users.noreply.github.com>
Date: Sat, 19 Sep 2026 03:43:24 -0700
Subject: [PATCH] fix(cache): reserve capacity after resize eviction

---
 litellm/caching/in_memory_cache.py            |  3 ++
 .../caching/test_in_memory_cache.py           | 42 +++++++++++++++++--
 2 files changed, 41 insertions(+), 4 deletions(-)

diff --git a/litellm/caching/in_memory_cache.py b/litellm/caching/in_memory_cache.py
index ce698f589ed..d8501508134 100644
--- a/litellm/caching/in_memory_cache.py
+++ b/litellm/caching/in_memory_cache.py
@@ -167,6 +167,9 @@ class InMemoryCache(BaseCache):
         if not self.check_value_size(value):
             return
 
+        if key not in self.cache_dict and len(self.cache_dict) >= self.max_size_in_memory:
+            self.evict_cache()
+
         self.cache_dict[key] = value
         if self.allow_ttl_override(key):  # if ttl is not set, set it to default ttl
             if "ttl" in kwargs and kwargs["ttl"] is not None:
diff --git a/tests/test_litellm/caching/test_in_memory_cache.py b/tests/test_litellm/caching/test_in_memory_cache.py
index 53e37c04c3e..054aacce0b0 100644
--- a/tests/test_litellm/caching/test_in_memory_cache.py
+++ b/tests/test_litellm/caching/test_in_memory_cache.py
@@ -4,7 +4,7 @@ import threading
 import time
 from concurrent.futures import ThreadPoolExecutor
 from typing import Final, Literal
-from unittest.mock import MagicMock, patch
+from unittest.mock import patch
 
 import httpx
 import pytest
@@ -13,6 +13,7 @@ from fastapi.testclient import TestClient
 
 from unittest.mock import AsyncMock
 
+from litellm.caching.dual_cache import DualCache
 from litellm.caching.in_memory_cache import InMemoryCache
 
 
@@ -22,6 +23,17 @@ class _SlowInt(int):
         return _SlowInt(int(self) + value)
 
 
+class _FakeClock:
+    def __init__(self) -> None:
+        self.now = 0.0
+
+    def __call__(self) -> float:
+        return self.now
+
+    def advance(self, seconds: float) -> None:
+        self.now += seconds
+
+
 def test_increment_cache_is_atomic_under_thread_concurrency():
     cache = InMemoryCache()
     seed = 1000
@@ -153,7 +165,7 @@ def test_in_memory_cache_max_size_with_ttl():
 
 @pytest.mark.parametrize("updated_key", ["first", "second"])
 def test_updating_full_cache_preserves_entries_and_expiration(updated_key: Literal["first", "second"]) -> None:
-    clock: Final = MagicMock(return_value=0.0)
+    clock: Final = _FakeClock()
     cache: Final = InMemoryCache(max_size_in_memory=2, clock=clock)
     cache.set_cache("first", "original-a", ttl=10)
     cache.set_cache("second", "original-b", ttl=30)
@@ -163,14 +175,36 @@ def test_updating_full_cache_preserves_entries_and_expiration(updated_key: Liter
     assert cache.get_cache("first") == ("updated" if updated_key == "first" else "original-a")
     assert cache.get_cache("second") == ("updated" if updated_key == "second" else "original-b")
 
-    clock.return_value = 11.0
+    clock.advance(11.0)
     assert cache.get_cache("first") is None
     assert cache.get_cache("second") == ("updated" if updated_key == "second" else "original-b")
 
-    clock.return_value = 31.0
+    clock.advance(20.0)
     assert cache.get_cache("second") is None
 
 
+@pytest.mark.parametrize(
+    ("updated_key", "expected_values"),
+    [
+        ("first", ("updated", None, "original-c")),
+        ("third", (None, "original-b", "updated")),
+    ],
+)
+def test_updating_cache_after_capacity_reduction_respects_limit(
+    updated_key: Literal["first", "third"], expected_values: tuple[str | None, str | None, str | None]
+) -> None:
+    cache: Final = InMemoryCache(max_size_in_memory=3, clock=_FakeClock())
+    cache.set_cache("first", "original-a", ttl=10)
+    cache.set_cache("second", "original-b", ttl=20)
+    cache.set_cache("third", "original-c", ttl=30)
+    DualCache(in_memory_cache=cache).update_in_memory_max_size(2)
+
+    cache.set_cache(updated_key, "updated", ttl=60)
+
+    assert len(cache.cache_dict) == 2
+    assert tuple(cache.get_cache(key) for key in ("first", "second", "third")) == expected_values
+
+
 def test_in_memory_cache_expired_items_evicted_first():
     """
     Test that expired items are evicted before non-expired items when cache is full.