perf: read shadow eval volume estimate from the daily rollup

The start endpoint counted the key's trailing 7-day requests directly
against LiteLLM_SpendLogs, which has no api_key index — on a busy proxy
that is a scan over every request in the window (potentially tens of
millions of rows) to answer one count, holding a multi-second query per
'Start shadow eval' click.

Read SUM(api_requests) from LiteLLM_DailyUserSpend instead: one indexed
row per key/day, the same table the usage dashboards already use for
this question. A regression test asserts the estimate queries the
rollup and never touches LiteLLM_SpendLogs.

Co-Authored-By: Claude <noreply@anthropic.com>

litellm/proxy/management_endpoints/auto_router_endpoints.py

@@ -51,6 +51,7 @@ from litellm.types.management_endpoints.auto_router_endpoints import (
 if TYPE_CHECKING:
     from fastapi import APIRouter, Depends, HTTPException, Query, status
 
+    from litellm.proxy.utils import PrismaClient
     from litellm.router import Router
 else:
     try:
@@ -453,6 +454,33 @@ _JUDGE_PROMPT_TOKENS_ESTIMATE: Final = 4000
 # The estimate projects from the key's request volume over this many trailing days.
 _ESTIMATE_LOOKBACK_DAYS: Final = 7
 
+_ESTIMATE_VOLUME_SQL: Final = """
+SELECT COALESCE(SUM(api_requests), 0)::bigint AS request_count
+FROM "LiteLLM_DailyUserSpend"
+WHERE api_key = $1 AND date >= $2
+"""
+
+
+class _EstimateVolumeRow(BaseModel):
+    request_count: int
+
+
+_ESTIMATE_VOLUME_ROWS: Final = TypeAdapter(list[_EstimateVolumeRow])
+
+
+async def _recent_request_volume(prisma_client: "PrismaClient", api_key_id: str) -> int:
+    """The key's request count over the estimate lookback window.
+
+    Read from the LiteLLM_DailyUserSpend rollup, never LiteLLM_SpendLogs: the raw log
+    table has no api_key index, so a per-key count there scans every request the proxy
+    served in the window. The rollup is a handful of indexed rows per key/day and is
+    how the usage dashboards answer the same question.
+    """
+    lookback_date: Final = (datetime.now(timezone.utc) - timedelta(days=_ESTIMATE_LOOKBACK_DAYS)).strftime("%Y-%m-%d")
+    raw_rows: Final = await prisma_client.db.query_raw(_ESTIMATE_VOLUME_SQL, api_key_id, lookback_date)
+    rows: Final = _ESTIMATE_VOLUME_ROWS.validate_python(raw_rows or [])
+    return rows[0].request_count if rows else 0
+
 
 def _require_admin_viewer(user_api_key_dict: UserAPIKeyAuth, action: str) -> None:
     if user_api_key_dict.user_role not in (
@@ -634,13 +662,7 @@ async def start_shadow_eval(
             detail=f"Key already has an active shadow eval job ({existing.id}). Stop it first.",
         )
 
-    lookback_start: Final = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(days=_ESTIMATE_LOOKBACK_DAYS)
-    recent_requests: Final = await prisma_client.db.litellm_spendlogs.count(
-        where={  # mutable-ok: Prisma filter
-            "api_key": data.api_key_id,
-            "startTime": {"gte": lookback_start},  # mutable-ok: Prisma filter
-        },
-    )
+    recent_requests: Final = await _recent_request_volume(prisma_client, data.api_key_id)
     sampled: Final = int(
         recent_requests * (data.duration_days / _ESTIMATE_LOOKBACK_DAYS) * data.shadow_percentage / 100.0
     )

tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py

@@ -574,7 +574,7 @@ class TestShadowEvalJobsAreTimeBound:
         router.adaptive_routers = {}
         router.quality_routers = {}
         prisma = MagicMock()
-        prisma.db.litellm_spendlogs.count = AsyncMock(return_value=recent_requests)
+        prisma.db.query_raw = AsyncMock(return_value=[{"request_count": recent_requests}])
         prisma.db.litellm_shadowevaljob.find_first = AsyncMock(return_value=None)
         created = MagicMock()
         created.id = "job-1"
@@ -583,6 +583,22 @@ class TestShadowEvalJobsAreTimeBound:
         monkeypatch.setattr(proxy_server, "prisma_client", prisma)
         return prisma
 
+    @pytest.mark.asyncio
+    async def test_estimate_reads_the_daily_rollup_not_the_raw_spend_log_table(self, monkeypatch: pytest.MonkeyPatch):
+        """LiteLLM_SpendLogs has no api_key index; a per-key count there scans every
+        request in the window. The estimate must come from LiteLLM_DailyUserSpend."""
+        from litellm.proxy.management_endpoints.auto_router_endpoints import start_shadow_eval
+
+        prisma = self._proxy_mocks(monkeypatch, recent_requests=700)
+
+        response = await start_shadow_eval(self._start_request(), ADMIN)
+
+        assert response.estimated_request_count == 70
+        prisma.db.litellm_spendlogs.count.assert_not_called()
+        sql = prisma.db.query_raw.call_args.args[0]
+        assert 'FROM "LiteLLM_DailyUserSpend"' in sql
+        assert "LiteLLM_SpendLogs" not in sql
+
     def test_duration_defaults_to_a_week_and_rejects_zero_and_over_a_month(self):
         assert self._start_request().duration_days == 7
         with pytest.raises(ValidationError):