doc

Author: akihironitta

src/lightning_app/components/auto_scaler.py

@@ -370,17 +370,17 @@ class AutoScaler(LightningFlow):
         # Example 2: Customizing the scaling logic
         class MyAutoScaler(L.app.components.AutoScaler):
             def scale(self, replicas: int, metrics: dict) -> int:
-                max_requests_per_work = self.max_batch_size
-                min_requests_per_work = 1
                 pending_requests_per_running_or_pending_work = metrics["pending_requests"] / (
                     replicas + metrics["pending_works"]
                 )
 
                 # upscale
+                max_requests_per_work = self.max_batch_size
                 if pending_requests_per_running_or_pending_work >= max_requests_per_work:
                     return replicas + 1
 
                 # downscale
+                min_requests_per_work = max_requests_per_work * 0.25
                 if pending_requests_per_running_or_pending_work < min_requests_per_work:
                     return replicas - 1