fix(spend): coerce Decimal and numeric-string aggregates in the benchmarks read path

litellm/proxy/spend_tracking/auto_router_benchmarks.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 from collections.abc import Mapping, Sequence
 from dataclasses import astuple, dataclass, fields
 from datetime import date, datetime, time, timedelta, timezone
+from decimal import Decimal
 from functools import reduce
 from typing import TYPE_CHECKING, Final
 
@@ -136,7 +137,16 @@ def _per(numerator: float, denominator: float) -> float:
 
 
 def _number(value: object) -> float:
-    return float(value) if isinstance(value, (int, float)) else 0.0
+    """Postgres aggregates surface through the driver as int, float, Decimal or a numeric
+    string depending on the column type (SUM over BIGINT and EXTRACT both yield NUMERIC)."""
+    if isinstance(value, (int, float, Decimal)):
+        return float(value)
+    if not isinstance(value, str):
+        return 0.0
+    try:
+        return float(value)
+    except ValueError:
+        return 0.0
 
 
 def _counters_from_row(row: Mapping[str, object]) -> _Counters:

tests/test_litellm/proxy/spend_tracking/test_auto_router_benchmarks.py

@@ -6,6 +6,7 @@ is exercised here without a database.
 
 import datetime as dt
 from dataclasses import fields
+from decimal import Decimal
 
 import pytest
 
@@ -233,6 +234,35 @@ class TestTotals:
         assert response.groups == ()
         assert response.totals.turns == 0
 
+    def test_decimal_and_string_aggregates_survive_the_driver(self):
+        """SUM over BIGINT and EXTRACT(EPOCH) yield NUMERIC, which the driver may hand
+        back as Decimal or a numeric string; neither may read as zero traffic."""
+        response = build_response(
+            rows=[
+                _row(
+                    "a",
+                    sessions=Decimal("2"),
+                    turns=Decimal("64"),
+                    total_tokens=Decimal("10000"),
+                    total_session_seconds="7200.5",
+                )
+            ],
+            start_date=dt.date(2026, 8, 1),
+            end_date=dt.date(2026, 8, 3),
+        )
+        assert response.totals.turns == 64
+        assert response.totals.total_tokens == 10_000
+        assert response.totals.avg_tokens_per_session == pytest.approx(5000.0)
+        assert response.totals.avg_session_seconds == pytest.approx(3600.25)
+
+    def test_a_malformed_aggregate_reads_as_zero_rather_than_failing_the_page(self):
+        response = build_response(
+            rows=[_row("a", total_tokens="not-a-number")],
+            start_date=dt.date(2026, 8, 1),
+            end_date=dt.date(2026, 8, 3),
+        )
+        assert response.totals.total_tokens == 0
+
     def test_group_identity_is_carried_through(self):
         response = build_response(
             rows=[_row("claude-auto")], start_date=dt.date(2026, 8, 1), end_date=dt.date(2026, 8, 3)