Merge pull request #39197 from BerriAI/litellm_e2e_reliability_retry_context_window

test(e2e): cover retry-on-timeout and the context-window fallback

tests/e2e/models.py

@@ -806,6 +806,7 @@ class LiteLLMParamsBody(BaseModel):
     mock_response: str | None = None
     timeout: float | None = None
     tpm: int | None = None
+    weight: int | None = None
 
 
 ModelMode = Literal["batch", "realtime", "image_generation"]
@@ -820,6 +821,7 @@ class ModelInfoBody(BaseModel):
     mode: ModelMode | None = None
     access_groups: list[str] | None = None
     team_id: str | None = None
+    allowed_fails_policy: dict[str, int] | None = None
 
 
 class ModelNewBody(BaseModel):

tests/e2e/router/reliability_support.py

@@ -19,6 +19,8 @@ from models import (
     ChatMessage,
     ChatResponse,
     LiteLLMParamsBody,
+    ModelInfoBody,
+    ModelNewBody,
     ReliabilityChatBody,
     RouterSettingsOverride,
 )
@@ -26,6 +28,18 @@ from models import (
 REAL_MODEL = "openai/gpt-5.5"
 REAL_KEY = "os.environ/OPENAI_API_KEY"
 
+# The smallest-context chat model OpenAI still serves (16385 tokens). A prompt
+# past that limit comes back as a real `context_length_exceeded` 400, which is
+# what litellm maps to ContextWindowExceededError.
+SMALL_CONTEXT_MODEL = "openai/gpt-3.5-turbo"
+SMALL_CONTEXT_LIMIT_TOKENS = 16385
+
+
+def oversized_prompt(marker: str) -> str:
+    """A prompt comfortably past SMALL_CONTEXT_MODEL's context limit, so the
+    provider refuses it on length rather than answering a truncated version."""
+    return f"{marker} " + ("token " * (SMALL_CONTEXT_LIMIT_TOKENS + 4000))
+
 
 def create_bad_base_deployment(proxy: ProxyClient, name: str) -> str:
     """Register a deployment pointing at an unreachable base, so every call to it
@@ -40,6 +54,38 @@ def create_timeout_deployment(proxy: ProxyClient, name: str) -> str:
     return proxy.create_model(name, LiteLLMParamsBody(model=REAL_MODEL, api_key=REAL_KEY, timeout=0.001))
 
 
+def create_small_context_deployment(proxy: ProxyClient, name: str) -> str:
+    """Register a deployment on the smallest-context model OpenAI still serves, so an
+    oversized prompt earns a real context-window refusal from the provider."""
+    return proxy.create_model(name, LiteLLMParamsBody(model=SMALL_CONTEXT_MODEL, api_key=REAL_KEY))
+
+
+def create_always_timing_out_deployment(proxy: ProxyClient, name: str) -> str:
+    """The always-picked half of a retry pair: a 1ms deadline the backend always
+    exceeds, all of the model group's shuffle weight, and a cooldown policy that
+    benches it on its first Timeout so the retry cannot land on it again."""
+    return proxy.register_model(
+        ModelNewBody(
+            model_name=name,
+            litellm_params=LiteLLMParamsBody(model=REAL_MODEL, api_key=REAL_KEY, timeout=0.001, weight=1),
+            model_info=ModelInfoBody(allowed_fails_policy={"TimeoutErrorAllowedFails": 0}),
+        )
+    )
+
+
+def create_zero_weight_backup_deployment(proxy: ProxyClient, name: str) -> str:
+    """The other half of a retry pair: healthy, but weight 0, so the weighted shuffle
+    never opens on it. It is reachable only once its sibling is benched and the
+    weighted pick falls through to a uniform one over what is left."""
+    return proxy.register_model(
+        ModelNewBody(
+            model_name=name,
+            litellm_params=LiteLLMParamsBody(model=REAL_MODEL, api_key=REAL_KEY, weight=0),
+            model_info=ModelInfoBody(),
+        )
+    )
+
+
 def chat_override(
     proxy: ProxyClient,
     key: str,

tests/e2e/router/test_reliability_fallbacks_e2e.py

@@ -9,6 +9,10 @@ in the x-litellm-attempted-fallbacks header. Empty content is accepted only when
 `finish_reason == "length"` and the response billed completion tokens, since
 gpt-5.5 counts reasoning against max_tokens and can consume the whole budget
 before emitting any text; a fallback that produced nothing at all still fails.
+
+The context-window case is a different reroute from a plain failure: the provider
+refuses the prompt on length, and `context_window_fallbacks` is the setting that
+reroutes it, not `fallbacks`.
 """
 
 from __future__ import annotations
@@ -25,8 +29,10 @@ from reliability_support import (
     completion_tokens_of,
     content_of,
     create_bad_base_deployment,
+    create_small_context_deployment,
     create_timeout_deployment,
     finish_reason_of,
+    oversized_prompt,
     reasoning_tokens_of,
 )
 
@@ -82,3 +88,17 @@ class TestReliabilityFallbacks:
             override=RouterSettingsOverride(fallbacks=[{primary: ["gpt-5.5"]}]),
         )
         _assert_served_by_fallback(resp)
+
+    @pytest.mark.covers("reliability.fallback.context_window.routes_to_fallback")
+    def test_context_window_routes_to_fallback(
+        self, client: ComplexityRouterClient, resources: ResourceManager, scoped_key: str
+    ) -> None:
+        primary = f"reliability-ctxfail-{unique_marker()}"
+        model_id = create_small_context_deployment(client.proxy, primary)
+        resources.defer(lambda: client.proxy.delete_model(model_id))
+
+        resp = chat_override(
+            client.proxy, scoped_key, primary, oversized_prompt(unique_marker()),
+            override=RouterSettingsOverride(context_window_fallbacks=[{primary: ["gpt-5.5"]}]),
+        )
+        _assert_served_by_fallback(resp)

tests/e2e/router/test_reliability_retries_e2e.py

@@ -0,0 +1,73 @@
+"""Live e2e: a request that fails on its first deployment is retried inside its own
+model group and still comes back a completion.
+
+The model group is a pair: an always-timing-out deployment that holds all of the
+group's shuffle weight, and a healthy backup at weight 0. The weighted pick always
+opens on the timing-out one, its first Timeout benches it (an
+`allowed_fails_policy` of `TimeoutErrorAllowedFails: 0`), and the retry falls
+through to the only deployment left. So the customer sees a completion and the
+proxy reports that it took a retry to get there, with no random first pick in the
+middle of it.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from complexity_router_client import ComplexityRouterClient
+from e2e_config import unique_marker
+from lifecycle import ResourceManager
+from models import RouterSettingsOverride
+from reliability_support import (
+    chat_override,
+    completion_tokens_of,
+    content_of,
+    create_always_timing_out_deployment,
+    create_zero_weight_backup_deployment,
+    finish_reason_of,
+)
+
+pytestmark = pytest.mark.e2e
+
+
+class TestReliabilityRetries:
+    @pytest.mark.covers("reliability.retry.timeout.succeeds_within_retries")
+    def test_timeout_on_first_deployment_succeeds_on_retry(
+        self, client: ComplexityRouterClient, resources: ResourceManager, scoped_key: str
+    ) -> None:
+        group = f"reliability-retry-{unique_marker()}"
+        timing_out = create_always_timing_out_deployment(client.proxy, group)
+        resources.defer(lambda: client.proxy.delete_model(timing_out))
+        backup = create_zero_weight_backup_deployment(client.proxy, group)
+        resources.defer(lambda: client.proxy.delete_model(backup))
+
+        resp = chat_override(
+            client.proxy,
+            scoped_key,
+            group,
+            f"say hi {unique_marker()}",
+            override=RouterSettingsOverride(num_retries=2),
+        )
+
+        assert resp.status_code == 200, (
+            f"the retry should have landed on the healthy backup, got {resp.status_code}: {resp.body[:300]}"
+        )
+
+        attempted = resp.headers.get("x-litellm-attempted-retries")
+        assert attempted is not None, "response is missing the x-litellm-attempted-retries header"
+        assert int(attempted) >= 1, (
+            f"x-litellm-attempted-retries is {attempted!r}; a 200 with no retry means the request never "
+            "opened on the timing-out deployment, so this proves nothing about retries"
+        )
+
+        content = content_of(resp)
+        finish_reason = finish_reason_of(resp)
+        completion_tokens = completion_tokens_of(resp) or 0
+        assert isinstance(content, str), (
+            f"the retry should have returned a completion body, got content {content!r} (body={resp.body[:300]})"
+        )
+        assert content or (finish_reason == "length" and completion_tokens > 0), (
+            f"the retry returned empty content with finish_reason={finish_reason!r}, "
+            f"completion_tokens={completion_tokens}; empty content is only acceptable when the budget "
+            f"was spent on non-visible reasoning (body={resp.body[:300]})"
+        )