From d0c6e30838912e203674bd29b05e504d9629d67d Mon Sep 17 00:00:00 2001
From: RosieOh <20172207@gm.hannam.ac.kr>
Date: Wed, 28 Oct 2026 12:03:28 +0900
Subject: [PATCH] refactor(proxy): keep the env re-probe task in a set instead
 of a module global

The strict ruff gate now rejects new global statements (PLW0603), so the
once-per-process re-probe is tracked in a set that drops a task when it
finishes. A test covers that a second miss doesn't start a second re-probe
while one is running, and that a finished one is replaced.

The coordination Redis test helper now also isolates saml_sso_cache, which
_attach_redis_usage_cache writes to, so these tests stop leaking a stand-in
Redis into it.
---
 litellm/proxy/proxy_server.py                 | 10 +++---
 ...test_proxy_server_endpoints_and_startup.py | 33 +++++++++++++++++++
 2 files changed, 39 insertions(+), 4 deletions(-)

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index bc9c6878d57..f7963e42299 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -5110,7 +5110,8 @@ def build_redis_usage_cache(redis_params: Mapping[str, object]) -> RedisCache:
 _build_redis_usage_cache: Final = build_redis_usage_cache
 
 
-_coordination_redis_env_retry_task: "asyncio.Task | None" = None
+# Holds the one background re-probe per process; a set, so scheduling never rebinds a module global.
+_coordination_redis_env_retry_tasks: Final[set["asyncio.Task[None]"]] = set()
 
 
 async def _probe_coordination_redis_from_environment() -> RedisCache | None:
@@ -5187,16 +5188,17 @@ async def _retry_coordination_redis_from_environment(enable_redis_auth_cache: bo
 
 def _schedule_coordination_redis_env_retry(enable_redis_auth_cache: bool) -> None:
     """Start the background re-probe, once per process."""
-    global _coordination_redis_env_retry_task  # rebind-ok: one retry task per process
-    if _coordination_redis_env_retry_task is not None and not _coordination_redis_env_retry_task.done():
+    if any(not task.done() for task in _coordination_redis_env_retry_tasks):
         return
     try:
         loop: Final = asyncio.get_running_loop()
     except RuntimeError:
         return
-    _coordination_redis_env_retry_task = loop.create_task(
+    task: Final = loop.create_task(
         _retry_coordination_redis_from_environment(enable_redis_auth_cache=enable_redis_auth_cache)
     )
+    _coordination_redis_env_retry_tasks.add(task)
+    task.add_done_callback(_coordination_redis_env_retry_tasks.discard)
 
 
 def environment_has_redis_connection_target() -> bool:
diff --git a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py
index faa0988bb32..83e9cc08809 100644
--- a/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py
+++ b/tests/unit/proxy/test_proxy_server_endpoints_and_startup.py
@@ -13169,6 +13169,7 @@ def _patched_coordination_redis_module_state(
         patch.object(proxy_server_module, "spend_counter_cache", spend_cache),
         patch.object(proxy_server_module, "user_api_key_cache", DualCache()),
         patch.object(proxy_server_module, "cli_sso_session_cache", DualCache()),
+        patch.object(proxy_server_module, "saml_sso_cache", DualCache()),
         patch.object(proxy_server_module, "llm_router", None),
         patch.object(proxy_server_module, "litellm_config_cache", config_cache),
         patch.object(proxy_server_module, "RedisCache", redis_cache_class),
@@ -13476,6 +13477,38 @@ async def test_env_fallback_does_not_schedule_a_retry_without_a_redis_target():
     assert scheduled.call_count == 0
 
 
+@pytest.mark.asyncio
+async def test_env_retry_is_scheduled_once_while_one_is_running():
+    """A config reload runs the env fallback again, so a second failed boot probe
+    must not start a second re-probe next to the one still running."""
+    release: Final = asyncio.Event()
+    started: Final[list[bool]] = []
+
+    async def retry(enable_redis_auth_cache: bool) -> None:
+        started.append(enable_redis_auth_cache)
+        await release.wait()
+
+    pending: Final[set[asyncio.Task[None]]] = set()
+    with (
+        patch.object(proxy_server_module, "_retry_coordination_redis_from_environment", retry),
+        patch.object(proxy_server_module, "_coordination_redis_env_retry_tasks", pending),
+    ):
+        proxy_server_module._schedule_coordination_redis_env_retry(enable_redis_auth_cache=False)
+        proxy_server_module._schedule_coordination_redis_env_retry(enable_redis_auth_cache=False)
+        assert len(pending) == 1
+
+        (first,) = tuple(pending)
+        release.set()
+        await first
+        assert not pending, "a finished re-probe is dropped so a later miss can start a new one"
+
+        proxy_server_module._schedule_coordination_redis_env_retry(enable_redis_auth_cache=True)
+        (second,) = tuple(pending)
+        await second
+
+    assert started == [False, True]
+
+
 @pytest.mark.asyncio
 async def test_retry_attaches_the_env_redis_once_it_answers():
     """The retry has to leave the pod in the state a healthy boot would: spend