feat(model-engine): add GcpPubSubQueueEndpointResourceDelegate for async endpoints on GCP

charts/model-engine/templates/_helpers.tpl

@@ -354,6 +354,7 @@ env:
   - name: LAUNCH_SERVICE_TEMPLATE_FOLDER
     value: "/workspace/model-engine/model_engine_server/infra/gateways/resources/templates"
   {{- $model_cache := default dict .Values.modelCache }}
+  {{- $gcp_cloud_provider := and .Values.config .Values.config.values .Values.config.values.infra (eq (.Values.config.values.infra.cloud_provider | default "") "gcp") }}
   - name: MODEL_CACHE_ENABLED
     value: {{ get $model_cache "enabled" | default false | quote }}
   - name: MODEL_CACHE_MOUNT_PATH
@@ -404,6 +405,14 @@ env:
   - name: SERVICEBUS_NAMESPACE
     value: {{ .Values.azure.servicebus_namespace }}
   {{- end }}
+  {{- if $gcp_cloud_provider }}
+  - name: GCP_PROJECT_ID
+    value: {{ (.Values.gcp).project_id | default "" | quote }}
+  - name: PUBSUB_TOPIC_PREFIX
+    value: {{ (.Values.gcp).pubsub_topic_prefix | default "" | quote }}
+  - name: PUBSUB_SUBSCRIPTION_PREFIX
+    value: {{ (.Values.gcp).pubsub_subscription_prefix | default "" | quote }}
+  {{- end }}
   {{- if eq .Values.context "circleci" }}
   - name: CIRCLECI
     value: "true"

charts/model-engine/templates/celery_autoscaler_stateful_set.yaml

@@ -5,12 +5,13 @@
 {{- $tag := .Values.tag }}
 {{- $message_broker := .Values.celeryBrokerType }}
 {{- $num_shards := .Values.celery_autoscaler.num_shards }}
+{{- $gcp_cloud_provider := and .Values.config .Values.config.values .Values.config.values.infra (eq (.Values.config.values.infra.cloud_provider | default "") "gcp") -}}
 {{- $broker_name := "redis-elasticache-message-broker-master" }}
 {{- if eq $message_broker "sqs" }}
 {{ $broker_name = "sqs-message-broker-master" }}
 {{- else if eq $message_broker "servicebus" }}
 {{ $broker_name = "servicebus-message-broker-master" }}
-{{- else if and .Values.config .Values.config.values .Values.config.values.infra (eq (.Values.config.values.infra.cloud_provider | default "") "gcp") }}
+{{- else if $gcp_cloud_provider }}
 {{ $broker_name = "redis-gcp-memorystore-message-broker-master" }}
 {{- end }}
 apiVersion: apps/v1
@@ -89,6 +90,14 @@ spec:
         - name: SERVICEBUS_NAMESPACE
           value: {{ .Values.azure.servicebus_namespace }}
         {{- end }}
+        {{- if $gcp_cloud_provider }}
+        - name: GCP_PROJECT_ID
+          value: {{ (.Values.gcp).project_id | default "" | quote }}
+        - name: PUBSUB_TOPIC_PREFIX
+          value: {{ (.Values.gcp).pubsub_topic_prefix | default "" | quote }}
+        - name: PUBSUB_SUBSCRIPTION_PREFIX
+          value: {{ (.Values.gcp).pubsub_subscription_prefix | default "" | quote }}
+        {{- end }}
         image: "{{ .Values.image.gatewayRepository }}:{{ $tag }}"
         imagePullPolicy: Always
         name: main

charts/model-engine/values.yaml

@@ -94,3 +94,9 @@ utilityImages:
 
 # Additional GPU tolerations for endpoint pods
 gpuTolerations: []
+
+# GCP configuration for GCP-based deployments
+gcp:
+  project_id: ""
+  pubsub_topic_prefix: "launch-endpoint-id-"
+  pubsub_subscription_prefix: "launch-endpoint-id-"

charts/model-engine/values_sample.yaml

@@ -438,3 +438,9 @@ recommendedHardware:
       gpu_type: nvidia-hopper-h100
       nodes_per_worker: 1
 #serviceBuilderQueue:
+
+# GCP configuration for GCP-based deployments
+gcp:
+  project_id: "your-gcp-project"
+  pubsub_topic_prefix: "launch-endpoint-id-"
+  pubsub_subscription_prefix: "launch-endpoint-id-"

clients/python/setup.py

@@ -5,5 +5,8 @@
     python_requires=">=3.8",
     version="0.0.0.beta45",
     packages=find_packages(),
+    # types-setuptools 82.0.0+ tightened package_data to _DictLike; the literal dict
+    # still works at runtime, only the new stub disagrees. Suppress at the call site
+    # rather than down-pinning the stub (which would mask real future tightenings).
     package_data={"llmengine": ["py.typed"]},  # type: ignore[arg-type]
 )

model-engine/model_engine_server/api/dependencies.py

@@ -95,6 +95,9 @@
 from model_engine_server.infra.gateways.resources.fake_queue_endpoint_resource_delegate import (
     FakeQueueEndpointResourceDelegate,
 )
+from model_engine_server.infra.gateways.resources.gcp_pubsub_queue_endpoint_resource_delegate import (
+    GcpPubSubQueueEndpointResourceDelegate,
+)
 from model_engine_server.infra.gateways.resources.live_endpoint_resource_gateway import (
     LiveEndpointResourceGateway,
 )
@@ -104,9 +107,6 @@
 from model_engine_server.infra.gateways.resources.queue_endpoint_resource_delegate import (
     QueueEndpointResourceDelegate,
 )
-from model_engine_server.infra.gateways.resources.redis_queue_endpoint_resource_delegate import (
-    RedisQueueEndpointResourceDelegate,
-)
 from model_engine_server.infra.gateways.resources.sqs_queue_endpoint_resource_delegate import (
     SQSQueueEndpointResourceDelegate,
 )
@@ -248,8 +248,17 @@ def _get_external_interfaces(
     elif infra_config().cloud_provider == "azure":
         queue_delegate = ASBQueueEndpointResourceDelegate()
     elif infra_config().cloud_provider == "gcp":
-        # GCP uses Redis (Memorystore) for Celery, so use Redis-based queue delegate
-        queue_delegate = RedisQueueEndpointResourceDelegate(redis_client=redis_client)
+        # Mirror the SQS_PROFILE env-first pattern: the Helm chart injects GCP_PROJECT_ID as a
+        # pod env var (from .Values.gcp.project_id), which is a different source than the YAML-
+        # rendered infra_service_config. Read the env first so the chart value reaches the delegate;
+        # the infra_config.gcp_project_id field handles setups that wire it via the config YAML.
+        gcp_project_id = os.getenv("GCP_PROJECT_ID") or infra_config().gcp_project_id
+        if not gcp_project_id:
+            raise ValueError(
+                "cloud_provider=gcp requires GCP_PROJECT_ID env var "
+                "(via .Values.gcp.project_id) or infra.gcp_project_id in the service config."
+            )
+        queue_delegate = GcpPubSubQueueEndpointResourceDelegate(project_id=gcp_project_id)
     else:
         queue_delegate = SQSQueueEndpointResourceDelegate(
             sqs_profile=os.getenv("SQS_PROFILE", hmi_config.sqs_profile)

model-engine/model_engine_server/core/config.py

@@ -54,6 +54,7 @@ class _InfraConfig:
     celery_enable_sha256: Optional[bool] = None
     docker_registry_type: Optional[str] = None
     debug_mode: Optional[bool] = None
+    gcp_project_id: Optional[str] = None
 
 
 @dataclass

model-engine/model_engine_server/entrypoints/k8s_cache.py

@@ -27,6 +27,9 @@
 from model_engine_server.infra.gateways.resources.fake_queue_endpoint_resource_delegate import (
     FakeQueueEndpointResourceDelegate,
 )
+from model_engine_server.infra.gateways.resources.gcp_pubsub_queue_endpoint_resource_delegate import (
+    GcpPubSubQueueEndpointResourceDelegate,
+)
 from model_engine_server.infra.gateways.resources.image_cache_gateway import ImageCacheGateway
 from model_engine_server.infra.gateways.resources.live_endpoint_resource_gateway import (
     LiveEndpointResourceGateway,
@@ -119,6 +122,16 @@ async def main(args: Any):
         queue_delegate = OnPremQueueEndpointResourceDelegate()
     elif infra_config().cloud_provider == "azure":
         queue_delegate = ASBQueueEndpointResourceDelegate()
+    elif infra_config().cloud_provider == "gcp":
+        # See dependencies.py for rationale: Helm injects GCP_PROJECT_ID as a pod env var;
+        # the infra_service_config YAML is a different source. Read the env first.
+        gcp_project_id = os.getenv("GCP_PROJECT_ID") or infra_config().gcp_project_id
+        if not gcp_project_id:
+            raise ValueError(
+                "cloud_provider=gcp requires GCP_PROJECT_ID env var "
+                "(via .Values.gcp.project_id) or infra.gcp_project_id in the service config."
+            )
+        queue_delegate = GcpPubSubQueueEndpointResourceDelegate(project_id=gcp_project_id)
     else:
         queue_delegate = SQSQueueEndpointResourceDelegate(
             sqs_profile=os.getenv("SQS_PROFILE", hmi_config.sqs_profile)

model-engine/model_engine_server/entrypoints/start_batch_job_orchestration.py

@@ -34,6 +34,9 @@
 from model_engine_server.infra.gateways.resources.fake_queue_endpoint_resource_delegate import (
     FakeQueueEndpointResourceDelegate,
 )
+from model_engine_server.infra.gateways.resources.gcp_pubsub_queue_endpoint_resource_delegate import (
+    GcpPubSubQueueEndpointResourceDelegate,
+)
 from model_engine_server.infra.gateways.resources.live_endpoint_resource_gateway import (
     LiveEndpointResourceGateway,
 )
@@ -90,6 +93,16 @@ async def run_batch_job(
         queue_delegate = OnPremQueueEndpointResourceDelegate()
     elif infra_config().cloud_provider == "azure":
         queue_delegate = ASBQueueEndpointResourceDelegate()
+    elif infra_config().cloud_provider == "gcp":
+        # See dependencies.py for rationale: Helm injects GCP_PROJECT_ID as a pod env var;
+        # the infra_service_config YAML is a different source. Read the env first.
+        gcp_project_id = os.getenv("GCP_PROJECT_ID") or infra_config().gcp_project_id
+        if not gcp_project_id:
+            raise ValueError(
+                "cloud_provider=gcp requires GCP_PROJECT_ID env var "
+                "(via .Values.gcp.project_id) or infra.gcp_project_id in the service config."
+            )
+        queue_delegate = GcpPubSubQueueEndpointResourceDelegate(project_id=gcp_project_id)
     else:
         queue_delegate = SQSQueueEndpointResourceDelegate(
             sqs_profile=os.getenv("SQS_PROFILE", hmi_config.sqs_profile)
@@ -110,6 +123,9 @@ async def run_batch_job(
     if infra_config().cloud_provider == "azure":
         inference_task_queue_gateway = servicebus_task_queue_gateway
         infra_task_queue_gateway = servicebus_task_queue_gateway
+    elif infra_config().cloud_provider == "gcp":
+        inference_task_queue_gateway = redis_task_queue_gateway
+        infra_task_queue_gateway = redis_task_queue_gateway
     elif infra_config().cloud_provider == "onprem" or infra_config().celery_broker_type_redis:
         # On-prem uses Redis-based task queues
         inference_task_queue_gateway = redis_task_queue_gateway

model-engine/model_engine_server/infra/gateways/resources/gcp_pubsub_queue_endpoint_resource_delegate.py

@@ -0,0 +1,151 @@
+from typing import Any, Dict, Optional
+
+from google.api_core import exceptions as gcp_exceptions
+from google.cloud import pubsub_v1
+from google.protobuf import field_mask_pb2
+from model_engine_server.core.loggers import logger_name, make_logger
+from model_engine_server.domain.exceptions import EndpointResourceInfraException
+from model_engine_server.infra.gateways.resources.queue_endpoint_resource_delegate import (
+    QueueEndpointResourceDelegate,
+    QueueInfo,
+)
+
+logger = make_logger(logger_name())
+
+GCP_PUBSUB_MAX_ACK_DEADLINE_SECONDS = 600  # Pub/Sub hard limit
+
+
+class GcpPubSubQueueEndpointResourceDelegate(QueueEndpointResourceDelegate):
+    """
+    Using GCP Pub/Sub (topic + subscription per endpoint).
+
+    topic_prefix and subscription_prefix control the GCP resource name prefix.
+    The logical queue_name returned to callers always uses the canonical
+    QueueEndpointResourceDelegate.endpoint_id_to_queue_name format, independent
+    of these prefixes.
+    """
+
+    def __init__(
+        self,
+        project_id: str,
+        topic_prefix: str = "launch-endpoint-id-",
+        subscription_prefix: str = "launch-endpoint-id-",
+    ) -> None:
+        if not project_id:
+            raise ValueError(
+                "GcpPubSubQueueEndpointResourceDelegate requires a non-empty project_id; "
+                "set infra.gcp_project_id in the service config."
+            )
+        self.project_id = project_id
+        self.topic_prefix = topic_prefix
+        self.subscription_prefix = subscription_prefix
+        # Lazily-initialized gRPC clients. Construction calls Google ADC which is
+        # unavailable in unit-test environments, so defer until first real use.
+        # The clients are then cached for the lifetime of the delegate.
+        self._publisher_client: Optional[pubsub_v1.PublisherClient] = None
+        self._subscriber_client: Optional[pubsub_v1.SubscriberClient] = None
+
+    @property
+    def _publisher(self) -> pubsub_v1.PublisherClient:
+        if self._publisher_client is None:
+            self._publisher_client = pubsub_v1.PublisherClient()
+        return self._publisher_client
+
+    @property
+    def _subscriber(self) -> pubsub_v1.SubscriberClient:
+        if self._subscriber_client is None:
+            self._subscriber_client = pubsub_v1.SubscriberClient()
+        return self._subscriber_client
+
+    def _topic_id(self, endpoint_id: str) -> str:
+        return f"{self.topic_prefix}{endpoint_id}"
+
+    def _subscription_id(self, endpoint_id: str) -> str:
+        return f"{self.subscription_prefix}{endpoint_id}"
+
+    async def create_queue_if_not_exists(
+        self,
+        endpoint_id: str,
+        endpoint_name: str,
+        endpoint_created_by: str,
+        endpoint_labels: Dict[str, Any],
+        queue_message_timeout_seconds: Optional[int] = None,
+    ) -> QueueInfo:
+        queue_name = QueueEndpointResourceDelegate.endpoint_id_to_queue_name(endpoint_id)
+        topic_path = f"projects/{self.project_id}/topics/{self._topic_id(endpoint_id)}"
+        subscription_path = (
+            f"projects/{self.project_id}/subscriptions/{self._subscription_id(endpoint_id)}"
+        )
+        ack_deadline = min(queue_message_timeout_seconds or 60, GCP_PUBSUB_MAX_ACK_DEADLINE_SECONDS)
+
+        try:
+            self._publisher.create_topic(name=topic_path)
+        except gcp_exceptions.AlreadyExists:
+            pass
+
+        try:
+            self._subscriber.create_subscription(
+                name=subscription_path,
+                topic=topic_path,
+                ack_deadline_seconds=ack_deadline,
+            )
+        except gcp_exceptions.AlreadyExists:
+            try:
+                self._subscriber.update_subscription(
+                    subscription=pubsub_v1.types.Subscription(
+                        name=subscription_path,
+                        ack_deadline_seconds=ack_deadline,
+                    ),
+                    update_mask=field_mask_pb2.FieldMask(paths=["ack_deadline_seconds"]),
+                )
+            except gcp_exceptions.GoogleAPIError as e:
+                logger.warning(
+                    f"Failed to update ack_deadline for Pub/Sub subscription {subscription_path}: {e}"
+                )
+
+        # Pub/Sub has no URL concept analogous to SQS queue URLs
+        return QueueInfo(queue_name, queue_url=None)
+
+    async def delete_queue(self, endpoint_id: str) -> None:
+        subscription_path = (
+            f"projects/{self.project_id}/subscriptions/{self._subscription_id(endpoint_id)}"
+        )
+        topic_path = f"projects/{self.project_id}/topics/{self._topic_id(endpoint_id)}"
+
+        # Always attempt BOTH deletions so a failure on one doesn't leave the other resource
+        # orphaned (Greptile P1). NotFound is silent. Other GoogleAPIErrors are collected and
+        # surfaced together at the end so callers see every cleanup failure, not just the first.
+        errors: list[tuple[str, str, gcp_exceptions.GoogleAPIError]] = []
+
+        try:
+            self._subscriber.delete_subscription(subscription=subscription_path)
+        except gcp_exceptions.NotFound:
+            logger.info(
+                f"Could not find Pub/Sub subscription {subscription_path} for endpoint {endpoint_id}"
+            )
+        except gcp_exceptions.GoogleAPIError as e:
+            errors.append(("subscription", subscription_path, e))
+
+        try:
+            self._publisher.delete_topic(topic=topic_path)
+        except gcp_exceptions.NotFound:
+            logger.info(f"Could not find Pub/Sub topic {topic_path} for endpoint {endpoint_id}")
+        except gcp_exceptions.GoogleAPIError as e:
+            errors.append(("topic", topic_path, e))
+
+        if errors:
+            details = "; ".join(
+                f"Failed to delete Pub/Sub {kind} {path}: {err}" for kind, path, err in errors
+            )
+            raise EndpointResourceInfraException(
+                f"Cleanup errors for endpoint {endpoint_id}: {details}"
+            ) from errors[0][2]
+
+    async def get_queue_attributes(self, endpoint_id: str) -> Dict[str, Any]:
+        queue_name = QueueEndpointResourceDelegate.endpoint_id_to_queue_name(endpoint_id)
+        return {
+            "name": queue_name,
+            # Pub/Sub does not expose a synchronous undelivered message count;
+            # real observability requires the Cloud Monitoring API as a separate concern.
+            "num_undelivered_messages": -1,
+        }

model-engine/model_engine_server/infra/gateways/resources/live_endpoint_resource_gateway.py

@@ -101,6 +101,14 @@ async def get_resources(
                 )
             elif "active_message_count" in sqs_attributes:  # from ASBQueueEndpointResourceDelegate
                 resources.num_queued_items = int(sqs_attributes["active_message_count"])
+            elif (
+                "num_undelivered_messages" in sqs_attributes
+            ):  # from GcpPubSubQueueEndpointResourceDelegate
+                # Pub/Sub returns -1 when num_undelivered_messages is not yet wired to Cloud Monitoring.
+                # Treat -1 as "unknown" and skip; downstream autoscaling expects non-negative counts.
+                gcp_count = int(sqs_attributes["num_undelivered_messages"])
+                if gcp_count >= 0:
+                    resources.num_queued_items = gcp_count
 
         return resources
 

model-engine/requirements.in

@@ -12,6 +12,12 @@ azure-storage-blob~=12.19.0
 # GCP dependencies
 gcloud-aio-storage~=9.6
 google-auth~=2.25.0
+google-cloud-pubsub>=2.18
+# google-cloud-pubsub transitively pulls opentelemetry-sdk, which flips
+# common/startup_tracing/correlation.py's OTEL_AVAILABLE to True. Once that's
+# True, tracer.py imports from opentelemetry-exporter-otlp-proto-grpc, which
+# isn't otherwise a dependency. Pin it explicitly so the import resolves.
+opentelemetry-exporter-otlp-proto-grpc
 google-cloud-artifact-registry~=1.21.0
 google-cloud-secret-manager>=2.24.0
 google-cloud-storage~=2.14.0