feat: Instrument pre-completion chat availability

src/mlpa/core/auth/authorize.py

@@ -13,6 +13,8 @@
     ServiceType,
 )
 from mlpa.core.config import env
+from mlpa.core.metrics import record_chat_availability_for
+from mlpa.core.prometheus_metrics import AvailabilityReason
 from mlpa.core.routers.appattest import app_attest_auth
 from mlpa.core.utils import extract_user_from_play_integrity_jwt, parse_app_attest_jwt
 
@@ -123,27 +125,55 @@ async def authorize_chat_request(
     )
 
     if not is_service_type_valid:
+        record_chat_availability_for(
+            AvailabilityReason.INVALID_SERVICE_TYPE_FOR_MODEL,
+            model=chat_request.model,
+            service_type=service_type.value,
+            purpose=purpose or "",
+        )
         raise HTTPException(
             status_code=400,
             detail=f"Invalid service-type value for model {chat_request.model}. Should be one of {env.forced_model_service_type_pairs.get(chat_request.model)}",
         )
 
-    return await _authorize_common_request(
-        request=request,
-        build_authorized_request=lambda user, purpose_value: AuthorizedChatRequest(
-            user=user,
+    try:
+        return await _authorize_common_request(
+            request=request,
+            build_authorized_request=lambda user, purpose_value: AuthorizedChatRequest(
+                user=user,
+                service_type=service_type.value,
+                purpose=purpose_value,
+                **chat_request.model_dump(exclude_unset=True),
+            ),
+            authorization=authorization,
+            service_type=service_type,
+            purpose=purpose,
+            x_dev_authorization=x_dev_authorization,
+            use_app_attest=use_app_attest,
+            use_qa_certificates=use_qa_certificates,
+            use_play_integrity=use_play_integrity,
+        )
+    except HTTPException as exc:
+        # Only record terminal HTTP failures from the shared auth call:
+        # - 401: expected or normalized auth rejection (bad creds, expired token, etc.)
+        # - 400: client error to the auth layer (invalid purpose, or malformed App
+        #        Attest base64 decoded in app_attest_auth before its try block)
+        # - anything else (e.g. App Attest's explicit 500): re-raised unrecorded;
+        #   auth-system-failure capture is left to a follow-on auth backend change
+        # Non-HTTPException errors are not caught here and propagate unrecorded.
+        if exc.status_code == 401:
+            reason = AvailabilityReason.AUTH_REJECTED
+        elif exc.status_code == 400:
+            reason = AvailabilityReason.INVALID_AUTH_REQUEST
+        else:
+            raise
+        record_chat_availability_for(
+            reason,
+            model=chat_request.model,
             service_type=service_type.value,
-            purpose=purpose_value,
-            **chat_request.model_dump(exclude_unset=True),
-        ),
-        authorization=authorization,
-        service_type=service_type,
-        purpose=purpose,
-        x_dev_authorization=x_dev_authorization,
-        use_app_attest=use_app_attest,
-        use_qa_certificates=use_qa_certificates,
-        use_play_integrity=use_play_integrity,
-    )
+            purpose=purpose or "",
+        )
+        raise
 
 
 async def authorize_search_request(

src/mlpa/core/completions.py

@@ -20,13 +20,15 @@
 from mlpa.core.logger import logger
 from mlpa.core.metrics import (
     extract_tool_names,
+    record_chat_availability,
     record_chat_request_rejection,
     record_completion_latency,
     record_completion_success,
     record_request_with_tools,
     record_ttft,
 )
 from mlpa.core.prometheus_metrics import (
+    AvailabilityReason,
     PrometheusRejectionReason,
     PrometheusResult,
 )
@@ -52,20 +54,31 @@ def _build_litellm_body(req: AuthorizedChatRequest, *, stream: bool) -> dict:
 async def get_or_create_user_for_completion(
     user_id: str, req: AuthorizedChatRequest | AuthorizedSearchRequest
 ):
-    """Wraps get_or_create_user and records a signup-cap rejection metric if applicable."""
+    """
+    Wraps get_or_create_user and records availability for chat requests:
+    - signup cap (403 + MAX_USERS_REACHED): excluded, alongside the existing rejection metric
+    - user-resolution server or system failure (status >= 500): failure
+    - search requests and non-signup-cap, non-5xx failures: not recorded
+    """
     try:
         return await get_or_create_user(user_id)
     except HTTPException as exc:
-        if (
-            exc.status_code == 403
-            and isinstance(exc.detail, dict)
-            and exc.detail.get("error") == ERROR_CODE_MAX_USERS_REACHED
-            and isinstance(req, AuthorizedChatRequest)
-        ):
-            record_chat_request_rejection(
-                req,
-                PrometheusRejectionReason.SIGNUP_CAP_EXCEEDED,
-            )
+        if isinstance(req, AuthorizedChatRequest):
+            if (
+                exc.status_code == 403
+                and isinstance(exc.detail, dict)
+                and exc.detail.get("error") == ERROR_CODE_MAX_USERS_REACHED
+            ):
+                record_chat_request_rejection(
+                    req,
+                    PrometheusRejectionReason.SIGNUP_CAP_EXCEEDED,
+                )
+                record_chat_availability(req, AvailabilityReason.SIGNUP_CAP_EXCEEDED)
+            elif exc.status_code >= 500:
+                # User-resolution server or system failure. Non-signup-cap 4xx errors
+                # are not recorded; a client-side 4xx should get its own classification
+                # rather than counting as an availability failure.
+                record_chat_availability(req, AvailabilityReason.PROVISIONING_FAILURE)
         raise
 
 
@@ -80,6 +93,7 @@ async def stream_completion(
     record_request_with_tools(authorized_chat_request)
     body = _build_litellm_body(authorized_chat_request, stream=True)
     result = PrometheusResult.ERROR
+    availability_reason = AvailabilityReason.UPSTREAM_ERROR
     is_first_token = True
     prompt_tokens = 0
     completion_tokens = 0
@@ -139,6 +153,7 @@ async def _read_next_chunk(
                     if match.log_message:
                         logger.warning(match.log_message)
                     record_chat_request_rejection(authorized_chat_request, match.reason)
+                    availability_reason = match.availability_reason()
                     yield f'data: {{"error": {match.error_code}}}\n\n'.encode()
                     return
 
@@ -237,6 +252,7 @@ async def _read_next_chunk(
                 return
 
             if not streaming_started:
+                availability_reason = AvailabilityReason.EMPTY_RESPONSE
                 yield raise_and_log(
                     RuntimeError("LiteLLM returned an empty response"),
                     True,
@@ -256,6 +272,7 @@ async def _read_next_chunk(
                 snapshot=litellm_routing_snapshot,
             )
             result = PrometheusResult.SUCCESS
+            availability_reason = AvailabilityReason.VALID_RESPONSE
     except (GeneratorExit, asyncio.CancelledError):
         # Client went away mid-stream: Starlette tears the generator down by
         # throwing GeneratorExit (or cancelling the task) at the paused
@@ -291,9 +308,12 @@ async def _read_next_chunk(
         if result == PrometheusResult.ERROR and disconnect_event.is_set():
             result = PrometheusResult.ABORT
             logger.info(_client_disconnected_msg)
+        if result == PrometheusResult.ABORT:
+            availability_reason = AvailabilityReason.CLIENT_DISCONNECT
         record_completion_latency(
             authorized_chat_request, result, time.perf_counter() - start_time
         )
+        record_chat_availability(authorized_chat_request, availability_reason)
 
 
 async def get_completion(authorized_chat_request: AuthorizedChatRequest):
@@ -304,6 +324,7 @@ async def get_completion(authorized_chat_request: AuthorizedChatRequest):
     record_request_with_tools(authorized_chat_request)
     body = _build_litellm_body(authorized_chat_request, stream=False)
     result = PrometheusResult.ERROR
+    availability_reason = AvailabilityReason.UPSTREAM_ERROR
     logger.debug(
         f"Starting a non-stream completion using {authorized_chat_request.model}, for user {authorized_chat_request.user}",
     )
@@ -326,6 +347,7 @@ async def get_completion(authorized_chat_request: AuthorizedChatRequest):
                 if match.log_message:
                     logger.warning(match.log_message)
                 record_chat_request_rejection(authorized_chat_request, match.reason)
+                availability_reason = match.availability_reason()
                 headers = (
                     {"Retry-After": match.retry_after} if match.retry_after else None
                 )
@@ -362,6 +384,7 @@ async def get_completion(authorized_chat_request: AuthorizedChatRequest):
             snapshot=litellm_routing_snapshot,
         )
         result = PrometheusResult.SUCCESS
+        availability_reason = AvailabilityReason.VALID_RESPONSE
         return data
     except HTTPException:
         raise
@@ -371,3 +394,4 @@ async def get_completion(authorized_chat_request: AuthorizedChatRequest):
         record_completion_latency(
             authorized_chat_request, result, time.perf_counter() - start_time
         )
+        record_chat_availability(authorized_chat_request, availability_reason)

src/mlpa/core/errors.py

@@ -9,7 +9,7 @@
     ERROR_CODE_REQUEST_TOO_LARGE,
     ERROR_CODE_UPSTREAM_RATE_LIMIT_EXCEEDED,
 )
-from mlpa.core.prometheus_metrics import PrometheusRejectionReason
+from mlpa.core.prometheus_metrics import AvailabilityReason, PrometheusRejectionReason
 from mlpa.core.utils import (
     is_context_window_error,
     is_invalid_model_name_error,
@@ -18,6 +18,15 @@
     is_rate_limit_error,
 )
 
+_REJECTION_TO_AVAILABILITY_REASON: dict[
+    PrometheusRejectionReason, AvailabilityReason
+] = {
+    PrometheusRejectionReason.BUDGET_EXCEEDED: AvailabilityReason.BUDGET_EXCEEDED,
+    PrometheusRejectionReason.PAYLOAD_TOO_LARGE: AvailabilityReason.PAYLOAD_TOO_LARGE,
+    PrometheusRejectionReason.INVALID_MODEL_NAME: AvailabilityReason.INVALID_MODEL_NAME,
+    PrometheusRejectionReason.INVALID_REQUEST: AvailabilityReason.INVALID_REQUEST,
+}
+
 
 @dataclass(frozen=True)
 class RejectionMatch:
@@ -27,6 +36,15 @@ class RejectionMatch:
     retry_after: str | None = None
     log_message: str = ""
 
+    def availability_reason(self) -> AvailabilityReason:
+        # SIGNUP_CAP_EXCEEDED is recorded pre-completion, not via classify_upstream_error,
+        # so it is not in the mapping below.
+        if self.reason == PrometheusRejectionReason.RATE_LIMITED:
+            if self.error_code == ERROR_CODE_UPSTREAM_RATE_LIMIT_EXCEEDED:
+                return AvailabilityReason.RATE_LIMITED_UPSTREAM
+            return AvailabilityReason.RATE_LIMITED_PLATFORM
+        return _REJECTION_TO_AVAILABILITY_REASON[self.reason]
+
 
 _RATE_LIMIT_REJECTION: dict[int, tuple[PrometheusRejectionReason, str, str]] = {
     ERROR_CODE_BUDGET_LIMIT_EXCEEDED: (

src/mlpa/core/metrics.py

@@ -6,9 +6,11 @@
     LitellmRoutingSnapshot,
 )
 from mlpa.core.prometheus_metrics import (
+    AvailabilityReason,
     PrometheusRejectionReason,
     PrometheusResult,
     TokenType,
+    availability_outcome_for,
     metrics,
 )
 
@@ -41,6 +43,33 @@ def record_search_request_rejection(
     metrics.search_request_rejections.labels(reason=reason, **_search_labels(req)).inc()
 
 
+def record_chat_availability_for(
+    reason: AvailabilityReason,
+    *,
+    model: str,
+    service_type: str,
+    purpose: str,
+) -> None:
+    metrics.chat_availability.labels(
+        outcome=availability_outcome_for(reason),
+        reason=reason,
+        model=model,
+        service_type=service_type,
+        purpose=purpose,
+    ).inc()
+
+
+def record_chat_availability(
+    req: AuthorizedChatRequest, reason: AvailabilityReason
+) -> None:
+    record_chat_availability_for(
+        reason,
+        model=req.model,
+        service_type=req.service_type,
+        purpose=req.purpose,
+    )
+
+
 def record_completion_latency(
     req: AuthorizedChatRequest,
     result: PrometheusResult,

src/mlpa/core/middleware/request_size.py

@@ -3,6 +3,8 @@
 
 from mlpa.core.config import ERROR_CODE_REQUEST_TOO_LARGE, env
 from mlpa.core.logger import logger
+from mlpa.core.metrics import record_chat_availability_for
+from mlpa.core.prometheus_metrics import AvailabilityReason
 
 
 async def check_request_size_middleware(request: Request, call_next):
@@ -19,6 +21,14 @@ async def check_request_size_middleware(request: Request, call_next):
                     logger.warning(
                         f"Request size {size} bytes exceeds maximum {env.MAX_REQUEST_SIZE_BYTES} bytes"
                     )
+                    # `model` is in the request body, which we don't read here.
+                    # We reject on the Content-Length header without parsing it.
+                    record_chat_availability_for(
+                        AvailabilityReason.PAYLOAD_TOO_LARGE,
+                        model="",
+                        service_type=request.headers.get("service-type") or "",
+                        purpose=request.headers.get("purpose") or "",
+                    )
                     return JSONResponse(
                         status_code=413,
                         content={"error": ERROR_CODE_REQUEST_TOO_LARGE},