From ccb6cd880aeedbf9f8bc0e63638c56a296f1f346 Mon Sep 17 00:00:00 2001
From: Alexsander Hamir <alexsanderhamirgomesbaptista@gmail.com>
Date: Wed, 10 Dec 2025 12:31:58 -0800
Subject: [PATCH] [Perf] - Cut memory leak in half (#17784)

Stop passing references to large objects to fire-and-forget tasks.
---
 litellm/proxy/db/db_spend_update_writer.py | 15 ++++++++-------
 litellm/proxy/utils.py                     |  6 ------
 2 files changed, 8 insertions(+), 13 deletions(-)

diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py
index 715a6ebd25d..97f5cb83439 100644
--- a/litellm/proxy/db/db_spend_update_writer.py
+++ b/litellm/proxy/db/db_spend_update_writer.py
@@ -6,6 +6,7 @@ Module responsible for
 """
 
 import asyncio
+import copy
 import json
 import os
 import random
@@ -162,14 +163,14 @@ class DBSpendUpdateWriter:
             asyncio.create_task(
                 self._update_tag_db(
                     response_cost=response_cost,
-                    request_tags=payload.get("request_tags"),
+                    request_tags=copy.deepcopy(payload.get("request_tags")),
                     prisma_client=prisma_client,
                 )
             )
 
             if disable_spend_logs is False:
                 await self._insert_spend_log_to_db(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     prisma_client=prisma_client,
                 )
             else:
@@ -179,34 +180,34 @@ class DBSpendUpdateWriter:
 
             asyncio.create_task(
                 self.add_spend_log_transaction_to_daily_user_transaction(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     prisma_client=prisma_client,
                 )
             )
 
             asyncio.create_task(
                 self.add_spend_log_transaction_to_daily_end_user_transaction(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     prisma_client=prisma_client,
                 )
             )
 
             asyncio.create_task(
                 self.add_spend_log_transaction_to_daily_team_transaction(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     prisma_client=prisma_client,
                 )
             )
             asyncio.create_task(
                 self.add_spend_log_transaction_to_daily_org_transaction(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     org_id=org_id,
                     prisma_client=prisma_client,
                 )
             )
             asyncio.create_task(
                 self.add_spend_log_transaction_to_daily_tag_transaction(
-                    payload=payload,
+                    payload=copy.deepcopy(payload),
                     prisma_client=prisma_client,
                 )
             )
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 7512ddc3688..e486bac310d 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -1,6 +1,5 @@
 import asyncio
 import copy
-import gc
 import hashlib
 import json
 import os
@@ -3379,7 +3378,6 @@ class ProxyUpdateSpend:
                             headers={"Content-Type": "application/json"},
                         )
                         del json_data
-                        gc.collect()
                         if response.status_code == 200:
                             prisma_client.spend_log_transactions = (
                                 prisma_client.spend_log_transactions[
@@ -3401,9 +3399,6 @@ class ProxyUpdateSpend:
                             )
                             # Explicitly clear batch memory
                             del batch, batch_with_dates
-                            # Only run gc every 5 batches to reduce overhead
-                            if j % (BATCH_SIZE * 5) == 0:
-                                gc.collect()
 
                         prisma_client.spend_log_transactions = (
                             prisma_client.spend_log_transactions[len(logs_to_process) :]
@@ -3429,7 +3424,6 @@ class ProxyUpdateSpend:
         finally:
             # Clean up logs_to_process after all processing is complete
             del logs_to_process
-            gc.collect()
 
     @staticmethod
     def disable_spend_updates() -> bool: