test(perf): give the direct vs bridge cases their own ceilings at 1.2x the lane median

tests/e2e/perf/cases/chat_completions_2rps.py

@@ -5,6 +5,8 @@ from __future__ import annotations
 
 from typing import Final
 
+from setup.performance_test_case import Limit
+
 from .direct_vs_bridge import direct_vs_bridge
 
 CASE: Final = direct_vs_bridge(
@@ -12,4 +14,25 @@ CASE: Final = direct_vs_bridge(
     "chat_completions",
     direct="gpt-6-luna",
     bridge="gpt-6-luna-responses",
+    limits=(
+        Limit("unexpected_status_count", max=3),
+        Limit("gateway_added_seconds", percentile=50, max=0.326),
+        Limit("gateway_added_seconds", percentile=99, max=2.23),
+        Limit("litellm_overhead_seconds", percentile=50, max=0.0372),
+        Limit("litellm_overhead_seconds", percentile=99, max=1.53),
+        Limit("litellm_response_seconds", percentile=50, max=1.05),
+        Limit("litellm_response_seconds", percentile=99, max=3.06),
+        Limit("prometheus_redis_latency_mean_seconds", max=0.00372),
+        Limit("postgres_active_sessions", max=0.00636),
+        Limit("postgres_slowest_statement_seconds", max=0.042),
+        Limit("postgres_peak_connections", max=24),
+        Limit("time_to_first_token_seconds", percentile=50, max=1.62),
+        Limit("time_to_first_token_seconds", percentile=99, max=3.66),
+        Limit("memory_growth_megabytes", max=600),
+        Limit("cpu_seconds_per_request", max=0.577),
+        Limit("postgres_calls_per_request", max=2.6),
+        Limit("postgres_max_statement_per_request", max=0.427),
+        Limit("postgres_max_background_statement_per_request", max=1.08),
+        Limit("redis_calls_per_request", max=15.0),
+    ),
 )

tests/e2e/perf/cases/direct_vs_bridge.py

@@ -1,14 +1,15 @@
 """The mixed case's traffic on one public API, half of it on a deployment the gateway serves on
 that provider's own API (direct) and half on a deployment it has to translate to another provider
-API (bridge). Same 2 rps, request count, shapes, stream share, callers, guardrails and limits as
+API (bridge). Same 2 rps, request count, shapes, stream share, callers and guardrails as
 `mixed_size_requests_2rps`, so a gap between the two halves and the mixed baseline is the
-translation's cost."""
+translation's cost. Each case carries its own limits, 1.2x the median of ten lane runs on every
+stable tag from v1.103.0 to v1.106.0, the same rule as the other perf cases."""
 
 from __future__ import annotations
 
 from dataclasses import replace
 
-from setup.performance_test_case import Endpoint, PerformanceTestCase, VirtualKey
+from setup.performance_test_case import Endpoint, Limit, PerformanceTestCase, VirtualKey
 
 from .mixed_size_requests_2rps import CASE as MIXED
 
@@ -17,7 +18,9 @@ def _scoped(key: VirtualKey, direct: str) -> VirtualKey:
     return replace(key, models=(direct,)) if key.models else key
 
 
-def direct_vs_bridge(name: str, endpoint: Endpoint, *, direct: str, bridge: str) -> PerformanceTestCase:
+def direct_vs_bridge(
+    name: str, endpoint: Endpoint, *, direct: str, bridge: str, limits: tuple[Limit, ...]
+) -> PerformanceTestCase:
     return replace(
         MIXED,
         name=name,
@@ -28,4 +31,5 @@ def direct_vs_bridge(name: str, endpoint: Endpoint, *, direct: str, bridge: str)
         virtual_keys=tuple(_scoped(key, direct) for key in MIXED.virtual_keys),
         model_groups=(direct, bridge),
         traffic=replace(MIXED.traffic, endpoints={endpoint: 1.0}),
+        limits=limits,
     )

tests/e2e/perf/cases/messages_2rps.py

@@ -6,6 +6,8 @@ from __future__ import annotations
 
 from typing import Final
 
+from setup.performance_test_case import Limit
+
 from .direct_vs_bridge import direct_vs_bridge
 
 CASE: Final = direct_vs_bridge(
@@ -13,4 +15,25 @@ CASE: Final = direct_vs_bridge(
     "messages",
     direct="claude-sonnet-5",
     bridge="claude-sonnet-5-bedrock-converse",
+    limits=(
+        Limit("unexpected_status_count", max=3),
+        Limit("gateway_added_seconds", percentile=50, max=0.296),
+        Limit("gateway_added_seconds", percentile=99, max=3.36),
+        Limit("litellm_overhead_seconds", percentile=50, max=0.214),
+        Limit("litellm_overhead_seconds", percentile=99, max=3.15),
+        Limit("litellm_response_seconds", percentile=50, max=1.18),
+        Limit("litellm_response_seconds", percentile=99, max=3.31),
+        Limit("prometheus_redis_latency_mean_seconds", max=0.00216),
+        Limit("postgres_active_sessions", max=0.00816),
+        Limit("postgres_slowest_statement_seconds", max=0.042),
+        Limit("postgres_peak_connections", max=24),
+        Limit("time_to_first_token_seconds", percentile=50, max=1.56),
+        Limit("time_to_first_token_seconds", percentile=99, max=5.07),
+        Limit("memory_growth_megabytes", max=600),
+        Limit("cpu_seconds_per_request", max=0.74),
+        Limit("postgres_calls_per_request", max=2.59),
+        Limit("postgres_max_statement_per_request", max=0.427),
+        Limit("postgres_max_background_statement_per_request", max=0.856),
+        Limit("redis_calls_per_request", max=18.1),
+    ),
 )

tests/e2e/perf/cases/responses_2rps.py

@@ -5,6 +5,8 @@ from __future__ import annotations
 
 from typing import Final
 
+from setup.performance_test_case import Limit
+
 from .direct_vs_bridge import direct_vs_bridge
 
 CASE: Final = direct_vs_bridge(
@@ -12,4 +14,25 @@ CASE: Final = direct_vs_bridge(
     "responses",
     direct="gpt-6-luna",
     bridge="claude-sonnet-5",
+    limits=(
+        Limit("unexpected_status_count", max=3),
+        Limit("gateway_added_seconds", percentile=50, max=0.24),
+        Limit("gateway_added_seconds", percentile=99, max=2.05),
+        Limit("litellm_overhead_seconds", percentile=50, max=0.153),
+        Limit("litellm_overhead_seconds", percentile=99, max=1.8),
+        Limit("litellm_response_seconds", percentile=50, max=1.11),
+        Limit("litellm_response_seconds", percentile=99, max=3.27),
+        Limit("prometheus_redis_latency_mean_seconds", max=0.00228),
+        Limit("postgres_active_sessions", max=0.011),
+        Limit("postgres_slowest_statement_seconds", max=0.042),
+        Limit("postgres_peak_connections", max=26),
+        Limit("time_to_first_token_seconds", percentile=50, max=1.43),
+        Limit("time_to_first_token_seconds", percentile=99, max=2.69),
+        Limit("memory_growth_megabytes", max=600),
+        Limit("cpu_seconds_per_request", max=0.382),
+        Limit("postgres_calls_per_request", max=2.59),
+        Limit("postgres_max_statement_per_request", max=0.426),
+        Limit("postgres_max_background_statement_per_request", max=1.51),
+        Limit("redis_calls_per_request", max=15.7),
+    ),
 )