From 9c0cb3a334de553ba8f4e8079b49f0301c6f62be Mon Sep 17 00:00:00 2001
From: Stanislav Hejny <stanislav.hejny@eulisa.local>
Date: Wed, 16 Jul 2025 18:04:09 +0100
Subject: [PATCH] refactor: update backpressure logic to handle availability
 correctly

Co-authored-by: aider (openrouter/openai/o3-mini-high) <aider@aider.chat>
---
 amqp/adapter/backpressure_handler.py | 45 +++++++++++++++++-----------
 1 file changed, 28 insertions(+), 17 deletions(-)

diff --git a/amqp/adapter/backpressure_handler.py b/amqp/adapter/backpressure_handler.py
index f566e52..1d52fc6 100644
--- a/amqp/adapter/backpressure_handler.py
+++ b/amqp/adapter/backpressure_handler.py
@@ -192,20 +192,32 @@ class BackpressureHandler:
 
     async def check_overload_condition(self):
         """
-        Check if the current parallel executions exceed the limit (OVERLOAD)
-        or if the service has been idle for too long (IDLE).
+        Check if the current availability is too low (OVERLOAD)
+        or if the service has been idle for too long with high availability (IDLE).
+        
+        Note: current_availability represents available capacity, not used capacity.
+        - Low availability (close to 0) means OVERLOAD
+        - High availability (close to maximum) with no activity means IDLE
         """
         current_time = time.time()
         last_event_delta = current_time - self.last_backpressure_event_time
+        
+        # If maximum is not set or invalid, use the highest seen value
+        if self.maximum_load <= 0 and self.current_load > 0:
+            self.maximum_load = self.current_load
+        elif self.current_load > self.maximum_load:
+            # Update maximum if we see a higher value
+            self.maximum_load = self.current_load
+        
         logging_info(
-            "Backpressure: Check conditions, current=%s, max=%s, last_activity=%s",
+            "Backpressure: Check conditions, current_availability=%s, max=%s, last_activity=%s",
             self.current_load,
             self.maximum_load,
             last_event_delta,
         )
-
-        # Check for OVERLOAD condition
-        if self.current_load >= round(0.8 * self.maximum_load):
+    
+        # Check for OVERLOAD condition - low availability (less than 20% of maximum)
+        if self.current_load <= round(0.2 * self.maximum_load):
             # Check if the last backpressure event was not an overload
             if self.last_backpressure_event != ScalingRequestAlert.OVERLOAD:
                 # Trigger the overload event
@@ -213,44 +225,43 @@ class BackpressureHandler:
                 # update / reset time-window so that the OVERLOAD is not sent too often
                 self.last_backpressure_event_time = current_time
                 self.last_backpressure_event = ScalingRequestAlert.OVERLOAD
-
-        # Check for IDLE condition
-        elif self.current_load == 0:
+    
+        # Check for IDLE condition - high availability (more than 80% of maximum) with no activity
+        elif self.current_load >= round(0.8 * self.maximum_load):
             idle_duration = self.config.backpressure.idle_duration
             # Check if service has been idle for longer than the configured duration
             if (
                 last_event_delta > idle_duration
                 and self.last_backpressure_event == ScalingRequestAlert.IDLE
             ):
-
                 logging_info(
                     "Backpressure: Service has been idle for %s seconds (threshold: %s seconds)",
                     last_event_delta,
                     idle_duration,
                 )
-
+    
                 # Trigger the idle event
                 await self._handle_backpressure_idle_event()
                 # update / reset time-window so that the IDLE is not sent too often
                 self.last_backpressure_event_time = current_time
             else:
-                # Don't send IDLE right away when the usage drops to zero / below threshold, but wait for a while
+                # Don't send IDLE right away when the availability increases, but wait for a while
                 self.last_backpressure_event = ScalingRequestAlert.IDLE
-
+    
         # If neither OVERLOAD nor IDLE, and it's time for an update, send UPDATE event
         elif last_event_delta > self.config.backpressure.time_window:
             _scaling_request: ScaleRequestV1 = ScaleRequestV1(
                 self.swarm_service_id,
                 self.swarm_task_id,
                 self.maximum_load,
-                self.maximum_load - self.current_load,
+                self.current_load,  # Current availability is passed directly
                 ScalingRequestAlert.UPDATE,
             )
             # Address the message to any adapter capable of supporting BACKPRESSURE request
             await self.publish_backpressure_request(_scaling_request)
             self.last_backpressure_event_time = current_time
             self.last_backpressure_event = ScalingRequestAlert.UPDATE
-
+    
         self.update_last_data_message_time()
 
     async def _backpressure_monitor(self):
@@ -353,7 +364,7 @@ class BackpressureHandler:
             self.swarm_service_id,
             self.swarm_task_id,
             self.maximum_load,
-            self.maximum_load - self.current_load,
+            self.current_load,  # Current availability is passed directly
             ScalingRequestAlert.OVERLOAD,
         )
         # Address the message to any management service instance capable of supporting BACKPRESSURE request
@@ -366,7 +377,7 @@ class BackpressureHandler:
             self.swarm_service_id,
             self.swarm_task_id,
             self.maximum_load,
-            self.maximum_load - self.current_load,
+            self.current_load,  # Current availability is passed directly
             ScalingRequestAlert.IDLE,
         )
         # Address the message to any adapter capable of supporting BACKPRESSURE request