test(perf): add chat_completions_2rps, responses_2rps and messages_2rps direct vs bridge cases (#47752)

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

tests/e2e/gateway/perf_ci_config.yml

@@ -46,6 +46,11 @@ model_list:
       model: openai/gpt-6-luna
       api_base: os.environ/E2E_PERF_MOCK_OPENAI_URL
       api_key: perf-mock-not-checked
+  - model_name: gpt-6-luna-responses
+    litellm_params:
+      model: openai/responses/gpt-6-luna
+      api_base: os.environ/E2E_PERF_MOCK_OPENAI_URL
+      api_key: perf-mock-not-checked
   - model_name: claude-sonnet-5
     litellm_params:
       model: anthropic/claude-sonnet-5
@@ -58,6 +63,13 @@ model_list:
       aws_region_name: us-east-1
       aws_access_key_id: perf-mock-not-checked
       aws_secret_access_key: perf-mock-not-checked
+  - model_name: claude-sonnet-5-bedrock-converse
+    litellm_params:
+      model: bedrock/converse/anthropic.claude-sonnet-5
+      aws_bedrock_runtime_endpoint: os.environ/E2E_PERF_MOCK_BEDROCK_URL
+      aws_region_name: us-east-1
+      aws_access_key_id: perf-mock-not-checked
+      aws_secret_access_key: perf-mock-not-checked
 guardrails:
   - guardrail_name: perf-pre
     litellm_params:

tests/e2e/perf/cases/chat_completions_2rps.py

@@ -0,0 +1,15 @@
+"""`/v1/chat/completions` at 2 rps: `gpt-6-luna` is served by the OpenAI mock's chat completions
+route, `gpt-6-luna-responses` (`openai/responses/gpt-6-luna`) is bridged to its Responses route."""
+
+from __future__ import annotations
+
+from typing import Final
+
+from .direct_vs_bridge import direct_vs_bridge
+
+CASE: Final = direct_vs_bridge(
+    "chat_completions_2rps",
+    "chat_completions",
+    direct="gpt-6-luna",
+    bridge="gpt-6-luna-responses",
+)

tests/e2e/perf/cases/direct_vs_bridge.py

@@ -0,0 +1,31 @@
+"""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
+`mixed_size_requests_2rps`, so a gap between the two halves and the mixed baseline is the
+translation's cost."""
+
+from __future__ import annotations
+
+from dataclasses import replace
+
+from setup.performance_test_case import Endpoint, PerformanceTestCase, VirtualKey
+
+from .mixed_size_requests_2rps import CASE as MIXED
+
+
+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:
+    return replace(
+        MIXED,
+        name=name,
+        teams=(
+            replace(MIXED.team("team-a"), models=(direct, bridge)),
+            replace(MIXED.team("team-b"), models=(bridge,)),
+        ),
+        virtual_keys=tuple(_scoped(key, direct) for key in MIXED.virtual_keys),
+        model_groups=(direct, bridge),
+        traffic=replace(MIXED.traffic, endpoints={endpoint: 1.0}),
+    )

tests/e2e/perf/cases/messages_2rps.py

@@ -0,0 +1,16 @@
+"""`/v1/messages` at 2 rps: `claude-sonnet-5` is served by the Anthropic mock's Messages route,
+`claude-sonnet-5-bedrock-converse` (`bedrock/converse/anthropic.claude-sonnet-5`) is bridged to the
+Bedrock mock's Converse route."""
+
+from __future__ import annotations
+
+from typing import Final
+
+from .direct_vs_bridge import direct_vs_bridge
+
+CASE: Final = direct_vs_bridge(
+    "messages_2rps",
+    "messages",
+    direct="claude-sonnet-5",
+    bridge="claude-sonnet-5-bedrock-converse",
+)

tests/e2e/perf/cases/responses_2rps.py

@@ -0,0 +1,15 @@
+"""`/v1/responses` at 2 rps: `gpt-6-luna` is served by the OpenAI mock's Responses route,
+`claude-sonnet-5` is bridged to the Anthropic mock's Messages route."""
+
+from __future__ import annotations
+
+from typing import Final
+
+from .direct_vs_bridge import direct_vs_bridge
+
+CASE: Final = direct_vs_bridge(
+    "responses_2rps",
+    "responses",
+    direct="gpt-6-luna",
+    bridge="claude-sonnet-5",
+)

tests/e2e/perf/setup/observe.py

@@ -565,6 +565,7 @@ def _diagnostics(observation: Observation) -> list[str]:
     return [
         "  diagnostics (not limits):",
         f"    bodies off reply {observation.bodies_off_reply()}",
+        f"    mock routes {_mock_routes(observation)}",
         f"    guardrail calls per request {len(observation.guardrail_calls) / count:.3f}",
         f"    guardrail missed {observation.guardrail_missed() if observation.guardrail_calls else 'n/a'}",
         f"    retried share {len(observation.retried_markers) / count:.4f}",
@@ -575,6 +576,12 @@ def _diagnostics(observation: Observation) -> list[str]:
     ]
 
 
+def _mock_routes(observation: Observation) -> str:
+    """Which provider route each request landed on, so a direct-vs-bridge case shows its split."""
+    counts: Final = Counter(f"{seen.provider}/{seen.route}" for seen in observation.mock_requests)
+    return ", ".join(f"{route} {count}" for route, count in sorted(counts.items())) or "n/a"
+
+
 def _datastore_lines(observation: Observation) -> list[str]:
     """What the per-request Postgres and Redis counts are made of, so a red run names the statement or
     command that grew. `plumbing` (BEGIN, COMMIT, session settings) and `unclassified` count in neither limit."""