fix(logging): leave oversized bodies out of the raw request log

With log_raw_request_response on, every request builds a curl string of the
full body and runs secret redaction over it on the event loop. For a 16 MB
body that blocks the proxy for seconds, so liveness probes time out and the
pod restarts

Bodies longer than MAX_RAW_REQUEST_BODY_LOG_CHARS (default 1,000,000) are
now replaced by a short placeholder in the curl string, which keeps the URL
and masked headers. raw_request_typed_dict still holds the full body

Fixes #49673

litellm/constants.py

@@ -2184,6 +2184,7 @@ MICROSOFT_USER_LAST_NAME_ATTRIBUTE: Final = str(os.getenv("MICROSOFT_USER_LAST_N
 # Maximum payload size (in bytes) to fully serialize for DEBUG logging.
 # Payloads larger than this are truncated to avoid multi-second json.dumps blocking the response.
 MAX_PAYLOAD_SIZE_FOR_DEBUG_LOG: Final = get_env_int("MAX_PAYLOAD_SIZE_FOR_DEBUG_LOG", 102400)  # 100 KB
+MAX_RAW_REQUEST_BODY_LOG_CHARS: Final = get_env_int("MAX_RAW_REQUEST_BODY_LOG_CHARS", 1_000_000)
 
 # Policy template enrichment
 MAX_COMPETITOR_NAMES: Final = get_env_int("MAX_COMPETITOR_NAMES", 100)

litellm/litellm_core_utils/litellm_logging.py

@@ -47,6 +47,7 @@ from litellm.constants import (
     INTERNAL_CALLBACK_NAMES,
     KUBERNETES_POD_ROUTING_KEY,
     LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL,
+    MAX_RAW_REQUEST_BODY_LOG_CHARS,
     PROVIDER_REQUEST_ID_HEADERS,
     REALTIME_METERED_USAGE_KEY,
     REDACTED_BY_LITELLM,
@@ -1830,7 +1831,10 @@ class Logging(LiteLLMLoggingBaseClass):
                     verbose_logger.debug("\x1b[92m%s\x1b[0m\n", curl_command)
 
     def _get_request_body(self, data: dict) -> str:
-        return str(data)
+        body: Final = str(data)
+        if len(body) > MAX_RAW_REQUEST_BODY_LOG_CHARS:
+            return f"<{len(body)} chars omitted, over MAX_RAW_REQUEST_BODY_LOG_CHARS={MAX_RAW_REQUEST_BODY_LOG_CHARS}>"
+        return body
 
     def _get_request_curl_command(self, api_base: str, headers: dict | None, additional_args: dict, data: dict) -> str:
         masked_api_base: Final = self._get_masked_api_base(api_base)

tests/unit/litellm_core_utils/test_litellm_logging.py

@@ -8662,6 +8662,30 @@ def test_pre_call_redacts_and_masks_raw_request(logging_obj):
     assert "key=*****" in raw_api_base
 
 
+def test_pre_call_omits_raw_request_body_over_size_limit(logging_obj, monkeypatch):
+    """Redacting a multi-MB raw request blocks the event loop, so an oversized body is left out of it."""
+    monkeypatch.setattr("litellm.litellm_core_utils.litellm_logging.MAX_RAW_REQUEST_BODY_LOG_CHARS", 100)
+    metadata = {"user_api_key_alias": "qa-key"}
+    logging_obj.model_call_details["litellm_params"] = {"metadata": metadata}
+    logging_obj.log_raw_request_response = True
+    body = {"messages": [{"role": "user", "content": "PRIVATE-PHRASE " * 20}]}
+
+    logging_obj.pre_call(
+        input="hi",
+        api_key="",
+        additional_args={
+            "api_base": "https://api.openai.com/v1/chat/completions",
+            "headers": {},
+            "complete_input_dict": body,
+        },
+    )
+
+    raw_request = metadata["raw_request"]
+    assert "PRIVATE-PHRASE" not in raw_request
+    assert "https://api.openai.com/v1/chat/completions" in raw_request
+    assert logging_obj.model_call_details["raw_request_typed_dict"]["raw_request_body"] == body
+
+
 _PRIVATE_RAW_REQUEST_ARGS: Final = {
     "api_base": "https://api.openai.com/v1/chat/completions",
     "headers": {},