Http connection reset (#25434)

* Possible fix for connection reset happening during read(), since payload sizes are large there can be a connection reset mid read.

Goal: Reproduce this minimally with aiohttp and write test cases

* Add debug logs and manually capture 104 connection reset errors, and retry. This is a id fix and doesnt fix root cause.

Root cause could be due to firewall or proxy resets or connection resets due as rate limiting.

* Simplify exception handling in HTTP request methods

Removed ReadError from exception handling in multiple request methods.

litellm/llms/custom_httpx/http_handler.py

@@ -113,6 +113,23 @@ def _prepare_request_data_and_content(
     return request_data, request_content
 
 
+def _is_connection_reset_error(e: Exception) -> bool:
+    error_text = getattr(e, "text", str(e))
+    error_response = getattr(e, "response", None)
+    if error_response and hasattr(error_response, "text"):
+        error_text = getattr(error_response, "text", error_text)
+
+    if isinstance(error_text, bytes):
+        error_text = error_text.decode("utf-8", errors="ignore")
+
+    normalized_error_text = str(error_text).lower()
+    return (
+        "errno 104" in normalized_error_text
+        or "connection reset by peer" in normalized_error_text
+        or "connection reset" in normalized_error_text
+    )
+
+
 # Cache for SSL contexts to avoid creating duplicate contexts with the same configuration
 # Key: tuple of (cafile, ssl_security_level, ssl_ecdh_curve)
 # Value: ssl.SSLContext
@@ -482,7 +499,14 @@ class AsyncHTTPHandler:
                     params=params,
                     headers=headers,
                     stream=stream,
+                    content=content,
                 )
+            except Exception as retry_error:
+                verbose_logger.debug(
+                    "httpx.ConnectError retry failed in AsyncHTTPHandler.post(): %s",
+                    str(retry_error),
+                )
+                raise
             finally:
                 await new_client.aclose()
         except httpx.TimeoutException as e:
@@ -512,6 +536,32 @@ class AsyncHTTPHandler:
 
             raise e
         except Exception as e:
+            if _is_connection_reset_error(e):
+                verbose_logger.debug(
+                    "Detected connection reset error in AsyncHTTPHandler.post(); retrying with a new client"
+                )
+                new_client = self.create_client(
+                    timeout=timeout, event_hooks=self.event_hooks
+                )
+                try:
+                    return await self.single_connection_post_request(
+                        url=url,
+                        client=new_client,
+                        data=data,
+                        json=json,
+                        params=params,
+                        headers=headers,
+                        stream=stream,
+                        content=content,
+                    )
+                except Exception as retry_error:
+                    verbose_logger.debug(
+                        "Connection reset retry failed in AsyncHTTPHandler.post(): %s",
+                        str(retry_error),
+                    )
+                    raise
+                finally:
+                    await new_client.aclose()
             raise e
 
     async def put(

tests/test_litellm/llms/custom_httpx/test_http_handler.py

@@ -3,7 +3,7 @@ import os
 import pathlib
 import ssl
 import sys
-from unittest.mock import MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import certifi
 import httpx
@@ -658,3 +658,43 @@ async def test_httpx_handler_uses_env_user_agent(monkeypatch):
         assert req.headers.get("User-Agent") == "Claude Code"
     finally:
         await handler.close()
+
+
+@pytest.mark.asyncio
+async def test_async_http_handler_retries_on_connection_reset_error_text():
+    handler = AsyncHTTPHandler()
+    retry_client = MagicMock()
+    retry_client.aclose = AsyncMock()
+    retry_response = MagicMock(spec=httpx.Response)
+
+    handler.client.send = AsyncMock(
+        side_effect=Exception("[Errno 104] Connection reset by peer")
+    )
+    handler.create_client = MagicMock(return_value=retry_client)
+    handler.single_connection_post_request = AsyncMock(return_value=retry_response)
+
+    try:
+        response = await handler.post(
+            url="https://example.com",
+            data={"hello": "world"},
+            headers={"x-test": "1"},
+            content="payload",
+        )
+
+        assert response is retry_response
+        handler.create_client.assert_called_once_with(
+            timeout=handler.timeout, event_hooks=handler.event_hooks
+        )
+        handler.single_connection_post_request.assert_awaited_once_with(
+            url="https://example.com",
+            client=retry_client,
+            data={"hello": "world"},
+            json=None,
+            params=None,
+            headers={"x-test": "1"},
+            stream=False,
+            content="payload",
+        )
+        retry_client.aclose.assert_awaited_once()
+    finally:
+        await handler.close()