harden uvloop configuration: raise UV_THREADPOOL_SIZE and add LITELLM_DISABLE_UVLOOP escape hatch

UV_THREADPOOL_SIZE defaults to 4 in libuv. Under concurrent DNS and filesystem
activity, the resolver threads can queue behind other work; a localhost lookup
that takes microseconds under idle conditions can stall for seconds when the
pool is saturated. Raising the default to 16 in all production Dockerfiles
gives the process 4x more headroom without significant memory overhead.

Adds LITELLM_DISABLE_UVLOOP env var: when set to "1", "true", or "yes",
_get_loop_type returns "asyncio" instead of "uvloop". This lets operators
switch event loops without rebuilding the image when they need a workaround
for uvloop-specific issues.

Dockerfile

@@ -81,7 +81,8 @@ RUN apk add --no-cache bash openssl tzdata nodejs npm python3 libsndfile && \
     { apk del --no-cache npm 2>/dev/null || true; }
 
 WORKDIR /app
-ENV PATH="/app/.venv/bin:${PATH}"
+ENV PATH="/app/.venv/bin:${PATH}" \
+    UV_THREADPOOL_SIZE=16
 
 COPY --from=builder /app /app
 # Prisma binaries live in $HOME/.cache (default prisma-python location),

docker/Dockerfile.database

@@ -88,7 +88,8 @@ RUN apk add --no-cache bash openssl tzdata nodejs npm python3 libsndfile && \
     { apk del --no-cache npm 2>/dev/null || true; }
 
 WORKDIR /app
-ENV PATH="/app/.venv/bin:${PATH}"
+ENV PATH="/app/.venv/bin:${PATH}" \
+    UV_THREADPOOL_SIZE=16
 
 COPY --from=builder /app /app
 # Prisma binaries live in $HOME/.cache (default prisma-python location),

docker/Dockerfile.non_root

@@ -107,7 +107,8 @@ ENV PATH="/app/.venv/bin:${PATH}" \
     PRISMA_SKIP_POSTINSTALL_GENERATE=1 \
     PRISMA_HIDE_UPDATE_MESSAGE=1 \
     PRISMA_ENGINES_CHECKSUM_IGNORE_MISSING=1 \
-    PRISMA_OFFLINE_MODE=true
+    PRISMA_OFFLINE_MODE=true \
+    UV_THREADPOOL_SIZE=16
 
 RUN mkdir -p /nonexistent /var/lib/litellm/assets /var/lib/litellm/ui && \
     chown -R nobody:nogroup /app /var/lib/litellm/ui /var/lib/litellm/assets /nonexistent && \

litellm/proxy/proxy_cli.py

@@ -479,9 +479,14 @@ class ProxyInitializationHelpers:
 
     @staticmethod
     def _get_loop_type():
-        """Helper function to determine the event loop type based on platform"""
         if sys.platform in ("win32", "cygwin", "cli"):
-            return None  # Let uvicorn choose the default loop on Windows
+            return None
+        if os.environ.get("LITELLM_DISABLE_UVLOOP", "").strip().lower() in (
+            "1",
+            "true",
+            "yes",
+        ):
+            return "asyncio"
         return "uvloop"
 
     @staticmethod

tests/test_litellm/proxy/test_proxy_cli.py

@@ -354,11 +354,29 @@ class TestProxyInitializationHelpers:
         assert ProxyInitializationHelpers._is_port_in_use(8000) is False
 
     def test_get_loop_type(self):
-        # Test on Windows
         with patch("sys.platform", "win32"):
             assert ProxyInitializationHelpers._get_loop_type() is None
 
-        # Test on Linux
+        with patch("sys.platform", "linux"):
+            assert ProxyInitializationHelpers._get_loop_type() == "uvloop"
+
+    @patch.dict(os.environ, {"LITELLM_DISABLE_UVLOOP": "1"})
+    def test_get_loop_type_disable_uvloop_numeric(self):
+        with patch("sys.platform", "linux"):
+            assert ProxyInitializationHelpers._get_loop_type() == "asyncio"
+
+    @patch.dict(os.environ, {"LITELLM_DISABLE_UVLOOP": "true"})
+    def test_get_loop_type_disable_uvloop_string(self):
+        with patch("sys.platform", "linux"):
+            assert ProxyInitializationHelpers._get_loop_type() == "asyncio"
+
+    @patch.dict(os.environ, {"LITELLM_DISABLE_UVLOOP": "yes"})
+    def test_get_loop_type_disable_uvloop_yes(self):
+        with patch("sys.platform", "linux"):
+            assert ProxyInitializationHelpers._get_loop_type() == "asyncio"
+
+    @patch.dict(os.environ, {"LITELLM_DISABLE_UVLOOP": "0"})
+    def test_get_loop_type_disable_uvloop_false(self):
         with patch("sys.platform", "linux"):
             assert ProxyInitializationHelpers._get_loop_type() == "uvloop"