[Serve] Short circuit autoscaling metrics

python/ray/serve/_private/metrics_utils.py

@@ -411,6 +411,9 @@ def merge_instantaneous_total(
     if not active_series:
         return []
 
+    if len(active_series) == 1:
+        return active_series[0]
+
     # True k-way merge: heap maintains exactly k elements (one per series)
     # Each element is (timestamp, replica_id, iterator)
     merge_heap = []

python/ray/serve/tests/unit/test_metrics_utils.py

@@ -4,10 +4,10 @@
 import pytest
 
 from ray._common.test_utils import async_wait_for_condition
+from ray.serve._private.common import TimeStampedValue
 from ray.serve._private.metrics_utils import (
     InMemoryMetricsStore,
     MetricsPusher,
-    TimeStampedValue,
     aggregate_timeseries,
     merge_instantaneous_total,
     merge_timeseries_dicts,
@@ -691,46 +691,57 @@ def test_merge_instantaneous_total_combine_same_timestamp(self):
 
     def test_merge_instantaneous_total_edge_cases_rounding(self):
         """Test edge cases for timestamp rounding and combination."""
-        # Test rounding edge cases
+        # Test rounding edge cases with multiple replicas (rounding only happens with 2+ replicas)
         series1 = [
             TimeStampedValue(1.004999, 5.0),  # Should round to 1.0
             TimeStampedValue(1.005000, 7.0),  # Should round to 1.0 (round half to even)
             TimeStampedValue(1.005001, 9.0),  # Should round to 1.01
         ]
+        series2 = [
+            TimeStampedValue(1.003, 2.0),  # Should round to 1.0
+        ]
 
-        result = merge_instantaneous_total([series1])
+        result = merge_instantaneous_total([series1, series2])
 
         # Should have two distinct rounded timestamps
         expected_timestamps = [1.0, 1.01]
         actual_timestamps = [point.timestamp for point in result]
         assert actual_timestamps == expected_timestamps
 
         # Values should reflect the changes
-        # Both 1.004999 and 1.005000 round to 1.0, so they get combined
-        # Order: 1.004999 (0->5), then 1.005000 (5->7) - final value at 1.0 is 7.0
-        # Then 1.005001 (7->9) rounds to 1.01 - value at 1.01 is 9.0
+        # At rounded timestamp 1.0:
+        # - series2 starts at 2 (t=1.003)
+        # - series1 starts at 5 (t=1.004999), total: 2+5=7
+        # - series1 changes to 7 (t=1.005000), total: 2+7=9
+        # At rounded timestamp 1.01:
+        # - series1 changes to 9 (t=1.005001), total: 2+9=11
         expected = [
-            TimeStampedValue(
-                1.0, 7.0
-            ),  # Final state after all changes that round to 1.0 (1.004999: 0->5, 1.005000: 5->7)
-            TimeStampedValue(1.01, 9.0),  # State after change at 1.005001 (7->9)
+            TimeStampedValue(1.0, 9.0),  # Final state at rounded timestamp 1.0
+            TimeStampedValue(1.01, 11.0),  # State after change at 1.005001
         ]
         assert_timeseries_equal(result, expected)
 
     def test_merge_instantaneous_total_no_changes_filtered(self):
         """Test that zero-change events are filtered even with rounding."""
+        # Use multiple replicas to trigger the merge logic (single replica returns as-is)
         series1 = [
             TimeStampedValue(1.001, 5.0),  # Rounds to 1.00
             TimeStampedValue(1.004, 5.0),  # Also rounds to 1.00, no change
             TimeStampedValue(2.000, 7.0),  # Rounds to 2.00, change
         ]
+        series2 = [
+            TimeStampedValue(1.002, 3.0),  # Rounds to 1.00
+            TimeStampedValue(2.001, 3.0),  # Rounds to 2.00, no change
+        ]
 
-        result = merge_instantaneous_total([series1])
+        result = merge_instantaneous_total([series1, series2])
 
-        # Should only include points where value actually changed
+        # Should only include points where total value actually changed
+        # At 1.00: series2 starts at 3, then series1 starts at 5, total = 8
+        # At 2.00: series1 changes to 7, total = 3+7 = 10
         expected = [
-            TimeStampedValue(1.00, 5.0),  # Initial value
-            TimeStampedValue(2.00, 7.0),  # Value change
+            TimeStampedValue(1.00, 8.0),  # Initial combined value
+            TimeStampedValue(2.00, 10.0),  # Value change
         ]
         assert_timeseries_equal(result, expected)