scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-24 00:52:24 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 90

fix(proxy/utils): extract spend-log patch to module-level helper with retry

Browse source
This commit is contained in:
Ishaan Jaffer 2026-04-22 19:55:46 -07:00
parent a9f699c683
commit 35a5b4dd80
No known key found for this signature in database
1 changed files with 73 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

76
litellm/proxy/utils.py
View file

@ -326,6 +326,50 @@ def _enrich_http_exception_with_guardrail_context(
detail.setdefault("guardrail_mode", event_hook)
async def _patch_eval_spend_log(
prisma_client: Any,
request_id: str,
eval_info: Any,
_retried: bool = False,
) -> None:
"""Patch eval_information into an existing spend log row.
The DB-logger callback that writes spend logs runs concurrently with
async_post_call_success_hook. If the row isn't committed yet, we sleep
500 ms and retry once — covering the common timing gap without busy-looping.
"""
import asyncio
import json as _json
try:
existing = await prisma_client.db.litellm_spendlogs.find_unique(
where={"request_id": request_id}
)
if existing is None:
if not _retried:
await asyncio.sleep(0.5)
await _patch_eval_spend_log(
prisma_client, request_id, eval_info, _retried=True
)
return
verbose_proxy_logger.warning(
f"[Evals] Spend log row {request_id} not found after retry; eval_information not persisted"
)
return
meta = existing.metadata or {}
if isinstance(meta, str):
meta = _json.loads(meta)
meta["eval_information"] = eval_info
await prisma_client.db.litellm_spendlogs.update(
where={"request_id": request_id},
data={"metadata": _json.dumps(meta)},
)
except Exception as _e:
verbose_proxy_logger.warning(
f"[Evals] Failed to patch spend log {request_id}: {_e}"
)
class ProxyLogging:
"""
Logging/Custom Handlers for proxy.
@ -2092,6 +2136,8 @@ class ProxyLogging:
############ Eval Gate [Beta] ##############################
############################################################
from litellm.proxy.proxy_server import prisma_client as _prisma_client
try:
_agent_id = (data.get("metadata") or {}).get("agent_id") or data.get(
"agent_id"
@ -2100,9 +2146,6 @@ class ProxyLogging:
from litellm.proxy.eval_management.eval_engine import (
run_evals_for_agent,
)
from litellm.proxy.proxy_server import (
prisma_client as _prisma_client,
)
if _prisma_client is not None:
response = await run_evals_for_agent(
@ -2112,10 +2155,37 @@ class ProxyLogging:
user_api_key_dict=user_api_key_dict,
prisma_client=_prisma_client,
)
# Patch the spend log — eval runs after the spend logger fires,
# so we update the row directly to make eval_information visible in Logs UI.
# For A2A calls, the spend log request_id is the JSON-RPC id (data["id"]),
# not the litellm_call_id UUID generated by common_request_processing.
_eval_info = (data.get("metadata") or {}).get("eval_information")
# For regular completions the spend log uses response.id (chatcmpl-XXX).
# For A2A calls it uses the JSON-RPC id stored in data["id"].
_request_id = (
getattr(response, "id", None)
or data.get("id")
or data.get("litellm_call_id")
)
if _eval_info and _request_id:
await _patch_eval_spend_log(
_prisma_client, _request_id, _eval_info
)
except Exception as e:
from fastapi import HTTPException
if isinstance(e, HTTPException):
# Even on block, patch the spend log so Logs UI shows the failed eval scores
_eval_info = (data.get("metadata") or {}).get("eval_information")
_request_id = (
getattr(response, "id", None)
or data.get("id")
or data.get("litellm_call_id")
)
if _eval_info and _request_id and _prisma_client is not None:
await _patch_eval_spend_log(
_prisma_client, _request_id, _eval_info
)
raise
verbose_proxy_logger.warning(
f"[Evals] Unexpected error in eval gate: {e}"