fix(alerts): stop detecting traffic drops (#51338)

litellm/models/performance_alert.py

@@ -17,8 +17,6 @@ SPEND_MULTIPLIER: Final = 3
 SPEND_MIN_INCREASE_USD: Final = 50
 TRAFFIC_SPIKE_MULTIPLIER: Final = 10
 TRAFFIC_SPIKE_MIN_INCREASE: Final = 1000
-TRAFFIC_DROP_RATIO: Final = 0.2
-TRAFFIC_DROP_MIN_BASELINE: Final = 500
 LATENCY_MULTIPLIER: Final = 3
 LATENCY_MIN_INCREASE_MS: Final = 500
 LATENCY_MIN_TIMED_REQUESTS: Final = 100
@@ -45,7 +43,6 @@ class ResourceWindow(BaseModel):
     base_timed: int = Field(ge=0, le=MAX_ALERT_ID)
     active_days: int = Field(ge=0, le=BASELINE_DAYS)
     timed_days: int = Field(ge=0, le=BASELINE_DAYS)
-    exists: bool
 
 
 class SweepState(BaseModel):
@@ -187,20 +184,14 @@ def _spend_spike(window: ResourceWindow) -> tuple[float, float] | None:
     return (window.spend, baseline) if triggered else None
 
 
-def _traffic_change(window: ResourceWindow) -> tuple[float, float] | None:
+def _traffic_spike(window: ResourceWindow) -> tuple[float, float] | None:
     baseline: Final = window.base_requests / BASELINE_DAYS
     spike: Final = (
         window.active_days >= MIN_ACTIVE_DAYS
         and window.requests >= TRAFFIC_SPIKE_MULTIPLIER * baseline
         and window.requests - baseline >= TRAFFIC_SPIKE_MIN_INCREASE
     )
-    drop: Final = (
-        window.exists
-        and window.active_days == BASELINE_DAYS
-        and baseline >= TRAFFIC_DROP_MIN_BASELINE
-        and window.requests < TRAFFIC_DROP_RATIO * baseline
-    )
-    return (window.requests, baseline) if spike or drop else None
+    return (window.requests, baseline) if spike else None
 
 
 def _slower_responses(window: ResourceWindow) -> tuple[float, float] | None:
@@ -231,7 +222,7 @@ def _more_failures(window: ResourceWindow) -> tuple[float, float] | None:
 def detect(scope: PerformanceScope, window: ResourceWindow, day: date) -> tuple[AlertDetection, ...]:
     checks: Final[tuple[tuple[AlertType, tuple[float, float] | None, int], ...]] = (
         ("spend", _spend_spike(window), window.requests),
-        ("traffic", _traffic_change(window), window.requests),
+        ("traffic", _traffic_spike(window), window.requests),
         ("latency", _slower_responses(window), window.timed),
         ("reliability", _more_failures(window), window.requests),
     )

litellm/repositories/performance_alert_repository.py

@@ -80,33 +80,18 @@ class _Source:
     resource_id: LiteralString
     requests: LiteralString
     filter: LiteralString
-    exists: LiteralString
 
 
 _SOURCES: Final[Mapping[PerformanceScope, _Source]] = {
-    "gateway": _Source('"LiteLLM_DailyModelUsage"', "'gateway'", "request_count", "", "true"),
-    "model": _Source('"LiteLLM_DailyModelUsage"', "model_group", "request_count", "", "true"),
-    "team": _Source(
-        '"LiteLLM_DailyTeamSpend"',
-        "team_id",
-        "api_requests",
-        "",
-        'EXISTS (SELECT 1 FROM "LiteLLM_TeamTable" t WHERE t.team_id = page.resource_id AND NOT t.blocked)',
-    ),
-    "user": _Source(
-        '"LiteLLM_DailyUserSpend"',
-        "user_id",
-        "api_requests",
-        "",
-        'EXISTS (SELECT 1 FROM "LiteLLM_UserTable" u WHERE u.user_id = page.resource_id)',
-    ),
+    "gateway": _Source('"LiteLLM_DailyModelUsage"', "'gateway'", "request_count", ""),
+    "model": _Source('"LiteLLM_DailyModelUsage"', "model_group", "request_count", ""),
+    "team": _Source('"LiteLLM_DailyTeamSpend"', "team_id", "api_requests", ""),
+    "user": _Source('"LiteLLM_DailyUserSpend"', "user_id", "api_requests", ""),
     "key": _Source(
         '"LiteLLM_DailyUserSpend"',
         "api_key",
         "api_requests",
         " AND api_key ~ '^[0-9a-f]{64}$'",
-        'EXISTS (SELECT 1 FROM "LiteLLM_VerificationToken" k WHERE k.token = page.resource_id'
-        " AND k.blocked IS NOT TRUE AND (k.expires IS NULL OR k.expires > now()))",
     ),
 }
 
@@ -127,8 +112,7 @@ def _page_sql(scope: PerformanceScope, wrapped: bool) -> LiteralString:
         if scope == "gateway"
         else f" GROUP BY {source.resource_id} ORDER BY {source.resource_id}"
     )
-    return f"""WITH page AS (
-    SELECT {source.resource_id} AS resource_id,
+    return f"""SELECT {source.resource_id} AS resource_id,
         COALESCE(SUM(spend) FILTER (WHERE date = $1), 0)::float8 AS spend,
         COALESCE(SUM(spend) FILTER (WHERE date < $1), 0)::float8 AS base_spend,
         COALESCE(SUM({source.requests}) FILTER (WHERE date = $1), 0)::bigint AS requests,
@@ -143,9 +127,7 @@ def _page_sql(scope: PerformanceScope, wrapped: bool) -> LiteralString:
         COUNT(DISTINCT date) FILTER (WHERE date < $1 AND timed_requests > 0) AS timed_days
     FROM {source.table}
     WHERE date BETWEEN $2 AND $1{where}{group}
-    LIMIT $3::int
-)
-SELECT page.*, {source.exists} AS exists FROM page ORDER BY resource_id"""
+    LIMIT $3::int"""
 
 
 def _next_id_sql(scope: PerformanceScope, wrapped: bool) -> LiteralString:

tests/integration/database/test_performance_alerts.py

@@ -157,6 +157,7 @@ async def test_a_full_sweep_saves_one_alert_per_real_change_and_nothing_for_norm
     write_rows('INSERT INTO "LiteLLM_UserTable" (user_id) VALUES (%s)', (f"{p}-user-stopped",))
     _seed("user", f"{p}-user-stopped", STEADY, None)
     _seed("user", f"{p}-user-deleted", STEADY, None)
+    _seed("user", f"{p}-user-busy", STEADY, Day(requests=20_000, timed=20_000, response_ms=2_000_000))
     _seed("key", fixture.key("slow"), STEADY, Day(response_ms=1_000_000))
     _seed("model", f"{p}-model", STEADY, Day(failed=100))
 
@@ -170,7 +171,7 @@ async def test_a_full_sweep_saves_one_alert_per_real_change_and_nothing_for_norm
     assert int(str(age["seconds"])) < 300
     assert fixture.saved() == {
         (f"{p}-team-spike", "spend"): ("team", "open"),
-        (f"{p}-user-stopped", "traffic"): ("user", "open"),
+        (f"{p}-user-busy", "traffic"): ("user", "open"),
         (fixture.key("slow"), "latency"): ("key", "open"),
         (f"{p}-model", "reliability"): ("model", "open"),
     }

tests/unit/models/test_performance_alert.py

@@ -5,6 +5,8 @@ import pytest
 
 from litellm.models.alert import AlertDetection, MetricEvidence
 from litellm.models.performance_alert import (
+    SCOPES,
+    PerformanceScope,
     ResourceWindow,
     SweepState,
     advance,
@@ -29,7 +31,6 @@ STEADY: Final = {
     "base_timed": 7000,
     "active_days": 7,
     "timed_days": 7,
-    "exists": True,
 }
 
 
@@ -45,6 +46,19 @@ def test_steady_usage_raises_nothing() -> None:
     assert _types() == ()
 
 
+@pytest.mark.parametrize("scope", SCOPES)
+@pytest.mark.parametrize("requests", (0, 100))
+def test_quiet_or_idle_resources_do_not_raise_traffic_alerts(scope: PerformanceScope, requests: int) -> None:
+    window: Final = _window(
+        resource_id="a" * 64 if scope == "key" else "resource-a",
+        requests=requests,
+        failed=0,
+        response_ms=requests * 100,
+        timed=requests,
+    )
+    assert detect(scope, window, DAY) == ()
+
+
 @pytest.mark.parametrize(
     ("changes", "expected"),
     (
@@ -57,11 +71,6 @@ def test_steady_usage_raises_nothing() -> None:
         ({"requests": 1100, "failed": 0, "response_ms": 110_000, "timed": 1100, "base_requests": 700}, ("traffic",)),
         ({"requests": 1099, "failed": 0, "response_ms": 109_900, "timed": 1099, "base_requests": 700}, ()),
         ({"requests": 50_000, "failed": 0, "active_days": 3}, ()),
-        ({"requests": 99, "failed": 0, "timed": 99, "response_ms": 9900, "base_requests": 3500}, ("traffic",)),
-        ({"requests": 100, "failed": 0, "timed": 100, "response_ms": 10_000, "base_requests": 3500}, ()),
-        ({"requests": 0, "failed": 0, "timed": 0, "response_ms": 0, "base_requests": 3500, "active_days": 6}, ()),
-        ({"requests": 0, "failed": 0, "timed": 0, "response_ms": 0, "base_requests": 3500, "exists": False}, ()),
-        ({"requests": 0, "failed": 0, "timed": 0, "response_ms": 0, "base_requests": 3493}, ()),
         ({"response_ms": 600_000}, ("latency",)),
         ({"response_ms": 599_000}, ()),
         ({"response_ms": 60_000_000, "timed": 99}, ()),

tests/unit/repositories/test_performance_alert_repository.py

@@ -112,7 +112,6 @@ def spend_window() -> Mapping[str, object]:
         "base_timed": 0,
         "active_days": 7,
         "timed_days": 0,
-        "exists": True,
     }