(ci/cd) final working load test / benchmark

.github/workflows/load_test.yml

@@ -23,8 +23,8 @@ jobs:
         with:
           LOCUSTFILE: ".github/workflows/locustfile.py"
           URL:  "https://litellm-database-docker-build-production.up.railway.app/"
-          USERS: "200"
-          RATE: "20"
+          USERS: "100"
+          RATE: "10"
           RUNTIME: "300s"
       - name: Process Load Test Stats
         run: |

.github/workflows/locustfile.py

@@ -6,7 +6,7 @@ import time
 class MyUser(HttpUser):
     wait_time = between(1, 5)
 
-    @task(20)
+    @task
     def chat_completion(self):
         headers = {
             "Content-Type": "application/json",
@@ -29,13 +29,13 @@ class MyUser(HttpUser):
 
         # Print or log the response if needed
 
-    @task(40)
+    @task(10)
     def health_readiness(self):
         start_time = time.time()
         response = self.client.get("health/readiness")
         response_time = time.time() - start_time
 
-    @task(40)
+    @task(10)
     def health_liveliness(self):
         start_time = time.time()
         response = self.client.get("health/liveliness")