diff --git a/litellm/constants.py b/litellm/constants.py index 334ef8d48a4..df6c77d533b 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1309,6 +1309,9 @@ MAX_SPENDLOG_ROWS_TO_QUERY = int( DEFAULT_SOFT_BUDGET = float( os.getenv("DEFAULT_SOFT_BUDGET", 50.0) ) # by default all litellm proxy keys have a soft budget of 50.0 +DEFAULT_CLIENT_DISCONNECT_CHECK_TIMEOUT_SECONDS = int( + os.getenv("DEFAULT_CLIENT_DISCONNECT_CHECK_TIMEOUT_SECONDS", 600) +) # 10 minutes timeout for client disconnect checking in proxy # makes it clear this is a rate limit error for a litellm virtual key RATE_LIMIT_ERROR_MESSAGE_FOR_VIRTUAL_KEY = "LiteLLM Virtual Key user_api_key_hash" diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index baa08537003..363fafcbd56 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -26,6 +26,7 @@ from litellm._logging import _redact_string, verbose_proxy_logger from litellm._uuid import uuid from litellm.constants import ( DD_TRACER_STREAMING_CHUNK_YIELD_RESOURCE, + DEFAULT_CLIENT_DISCONNECT_CHECK_TIMEOUT_SECONDS, DEFAULT_MAX_RECURSE_DEPTH, LITELLM_DETAILED_TIMING, MAX_PAYLOAD_SIZE_FOR_DEBUG_LOG, @@ -535,6 +536,39 @@ def _has_attribute_error_in_chain(exc: Exception) -> bool: return False +async def _check_request_disconnection( + request: Request, + llm_api_call_task, + disconnect_event: asyncio.Event, +): + """ + Asynchronously checks if the request is disconnected at regular intervals. + If the request is disconnected: + - cancels the litellm.router task (effective for non-streaming requests) + - sets disconnect_event so the caller can distinguish a client disconnect + from other sources of CancelledError (e.g. server shutdown) + + Parameters: + - request: The request object to check for disconnection. + - llm_api_call_task: The asyncio gather future to cancel on disconnect. + - disconnect_event: Event set when the client disconnects. + Returns: + - None + """ + + # only run this function for configured timeout -> if these don't get cancelled -> we don't want the server to have many while loops + start_time = time.time() + while time.time() - start_time < DEFAULT_CLIENT_DISCONNECT_CHECK_TIMEOUT_SECONDS: + await asyncio.sleep(1) + message = await request.receive() + if message.get("type") == "http.disconnect": + # cancel the LLM API Call task if any passed - this is passed from individual providers + # Example OpenAI, Azure, VertexAI etc + llm_api_call_task.cancel() + disconnect_event.set() + return + + class ProxyBaseLLMRequestProcessing: def __init__(self, data: dict): self.data = data @@ -1185,12 +1219,31 @@ class ProxyBaseLLMRequestProcessing: ) tasks.append(llm_call) - # wait for call to end llm_responses = asyncio.gather( *tasks ) # run the moderation check in parallel to the actual llm api call - responses = await llm_responses + # Execute the task to detect disconnection + disconnect_event = asyncio.Event() + disconnect_task = asyncio.create_task( + _check_request_disconnection(request, llm_responses, disconnect_event) + ) + + try: + # wait for call to end + # Note: for streaming requests llm_responses resolves quickly once the + # upstream connection is established; the ASGI transport layer handles + # cancellation of the upstream when the client disconnects mid-stream. + responses = await llm_responses + disconnect_task.cancel() + except asyncio.CancelledError: + disconnect_task.cancel() + if disconnect_event.is_set(): + raise HTTPException( + status_code=499, + detail="Client disconnected the request", + ) + raise response = responses[1] @@ -1727,9 +1780,13 @@ class ProxyBaseLLMRequestProcessing: version: Optional[str] = None, ): """Raises ProxyException (OpenAI API compatible) if an exception is raised""" - verbose_proxy_logger.exception( - f"litellm.proxy.proxy_server._handle_llm_api_exception(): Exception occured - {str(e)}" - ) + if isinstance(e, HTTPException) and e.status_code == 499: + verbose_proxy_logger.info("Client disconnected the request (499)") + raise e + else: + verbose_proxy_logger.exception( + f"litellm.proxy.proxy_server._handle_llm_api_exception(): Exception occured - {str(e)}" + ) # Allow callbacks to transform the error response transformed_exception = await proxy_logging_obj.post_call_failure_hook( user_api_key_dict=user_api_key_dict, diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 3db13e8ada6..cb6130420d0 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -1773,34 +1773,6 @@ db_writer_client: Optional[AsyncHTTPHandler] = None ### logger ### -async def check_request_disconnection(request: Request, llm_api_call_task): - """ - Asynchronously checks if the request is disconnected at regular intervals. - If the request is disconnected - - cancel the litellm.router task - - raises an HTTPException with status code 499 and detail "Client disconnected the request". - - Parameters: - - request: Request: The request object to check for disconnection. - Returns: - - None - """ - - # only run this function for 10 mins -> if these don't get cancelled -> we don't want the server to have many while loops - start_time = time.time() - while time.time() - start_time < 600: - await asyncio.sleep(1) - if await request.is_disconnected(): - # cancel the LLM API Call task if any passed - this is passed from individual providers - # Example OpenAI, Azure, VertexAI etc - llm_api_call_task.cancel() - - raise HTTPException( - status_code=499, - detail="Client disconnected the request", - ) - - def _resolve_typed_dict_type(typ): """Resolve the actual TypedDict class from a potentially wrapped type.""" from typing_extensions import _TypedDictMeta # type: ignore diff --git a/tests/test_litellm/proxy/test_client_disconnection.py b/tests/test_litellm/proxy/test_client_disconnection.py new file mode 100644 index 00000000000..bd0b3d5042e --- /dev/null +++ b/tests/test_litellm/proxy/test_client_disconnection.py @@ -0,0 +1,56 @@ +""" +Test client disconnection detection functionality. +""" + +import asyncio +import pytest +from unittest.mock import AsyncMock, MagicMock, patch + +from litellm.proxy.common_request_processing import _check_request_disconnection + + +@pytest.mark.asyncio +async def test_check_request_disconnection_with_disconnect(): + """Test that _check_request_disconnection cancels task and sets event when client disconnects.""" + mock_request = AsyncMock() + mock_request.receive.side_effect = [ + {"type": "http.request"}, # First call + {"type": "http.disconnect"}, # Second call - disconnect + ] + + mock_llm_task = MagicMock() # sync mock so .cancel() doesn't return a coroutine + disconnect_event = asyncio.Event() + + with patch( + "litellm.proxy.common_request_processing.asyncio.sleep", new_callable=AsyncMock + ): + await _check_request_disconnection( + mock_request, mock_llm_task, disconnect_event + ) + + mock_llm_task.cancel.assert_called_once() + assert disconnect_event.is_set() + + +@pytest.mark.asyncio +async def test_check_request_disconnection_no_disconnect(): + """Test that _check_request_disconnection does not cancel task during normal operation.""" + mock_request = AsyncMock() + mock_request.receive.return_value = {"type": "http.request"} + + mock_llm_task = MagicMock() # sync mock so .cancel() doesn't return a coroutine + disconnect_event = asyncio.Event() + + task = asyncio.create_task( + _check_request_disconnection(mock_request, mock_llm_task, disconnect_event) + ) + await asyncio.sleep(0.1) # Let it run briefly + task.cancel() + + try: + await task + except asyncio.CancelledError: + pass + + mock_llm_task.cancel.assert_not_called() + assert not disconnect_event.is_set()