Ensure web.Application is instrumented, and ensure the latest aiohttp is installed in containers

comfy/api_server/routes/internal/internal_routes.py

@@ -70,6 +70,12 @@ class InternalRoutes:
 
     def get_app(self):
         if self._app is None:
+            # Manually apply instrumentation to ensure web.Application is instrumented
+            from opentelemetry.instrumentation.aiohttp_server import AioHttpServerInstrumentor
+            instrumentor = AioHttpServerInstrumentor()
+            if not instrumentor.is_instrumented_by_opentelemetry:
+                instrumentor.instrument()
+
             self._app = web.Application()
             self.setup_routes()
             self._app.add_routes(self.routes)

comfy/cmd/main_pre.py

@@ -15,7 +15,6 @@ import shutil
 import warnings
 
 import fsspec
-from opentelemetry.instrumentation.urllib3 import URLLib3Instrumentor
 
 from .. import options
 from ..app import logger
@@ -130,6 +129,8 @@ def _create_tracer():
     from opentelemetry.processor.baggage import BaggageSpanProcessor, ALLOW_ALL_BAGGAGE_KEYS
     from opentelemetry.instrumentation.aiohttp_server import AioHttpServerInstrumentor
     from opentelemetry.instrumentation.aiohttp_client import AioHttpClientInstrumentor
+    from opentelemetry.instrumentation.urllib3 import URLLib3Instrumentor
+
     from ..tracing_compatibility import ProgressSpanSampler
     from ..tracing_compatibility import patch_spanbuilder_set_channel
 
@@ -163,6 +164,8 @@ def _create_tracer():
 
 
     provider.add_span_processor(BaggageSpanProcessor(ALLOW_ALL_BAGGAGE_KEYS))
+    trace.set_tracer_provider(provider)
+
     # makes this behave better as a library
     return trace.get_tracer(args.otel_service_name, tracer_provider=provider)
 

comfy/cmd/server.py

@@ -252,6 +252,14 @@ class PromptServer(ExecutorToClientProgress):
             middlewares.append(create_origin_only_middleware())
 
         max_upload_size = round(args.max_upload_size * 1024 * 1024)
+
+        # Manually apply instrumentation to ensure web.Application is instrumented
+        # This works around the import caching bug in opentelemetry-instrumentation-aiohttp-server
+        from opentelemetry.instrumentation.aiohttp_server import AioHttpServerInstrumentor
+        instrumentor = AioHttpServerInstrumentor()
+        if not instrumentor.is_instrumented_by_opentelemetry:
+            instrumentor.instrument()
+
         self.app: web.Application = web.Application(client_max_size=max_upload_size,
                                                     handler_args={'max_field_size': 16380},
                                                     middlewares=middlewares)

comfy/distributed/distributed_prompt_worker.py

@@ -53,6 +53,12 @@ class DistributedPromptWorker:
             return web.Response(text="UNHEALTHY: RabbitMQ connection is not healthy", status=503)
 
     async def _start_health_check_server(self):
+        # Manually apply instrumentation to ensure web.Application is instrumented
+        from opentelemetry.instrumentation.aiohttp_server import AioHttpServerInstrumentor
+        instrumentor = AioHttpServerInstrumentor()
+        if not instrumentor.is_instrumented_by_opentelemetry:
+            instrumentor.instrument()
+
         app = web.Application()
         app.router.add_get('/health', self._health_check)
 

pyproject.toml

@@ -33,8 +33,8 @@ dependencies = [
     "peft>=0.10.0",
     "torchinfo",
     "safetensors>=0.4.2",
-    "aiohttp>=3.11.8",
-    "yarl>=1.9.4",
+    "aiohttp>=3.13.2",
+    "yarl>=1.22.0",
     "accelerate>=0.25.0",
     "pyyaml>=6.0",
     "scikit-image>=0.20.0",
@@ -119,7 +119,8 @@ dependencies = [
     "requests_cache",
     "universal_pathlib",
     # yanked propcache is omitted
-    "propcache!=0.4.0",
+    # we want this upgraded
+    "propcache>=0.4.1",
 ]
 
 [build-system]