[BugFix] fix grpc failure when tracing init before workers forked (#6732)

* [fix] fix grpc failure when tracing init before workers forked

* [fix] change default exporter to http

* [fix] fix test_trace

fastdeploy/entrypoints/openai/api_server.py

@@ -84,7 +84,7 @@ from fastdeploy.utils import (
     retrive_model_from_server,
 )
 
-tracing.process_tracing_init()
+_tracing_inited = False
 
 parser = make_arg_parser(FlexibleArgumentParser())
 args = parser.parse_args()
@@ -171,8 +171,15 @@ async def lifespan(app: FastAPI):
     async context manager for FastAPI lifespan
     """
     global engine_args
+    global _tracing_inited
     import logging
 
+    # Initialize tracing in worker lifecycle instead of module import time.
+    # This avoids creating grpc/cygrpc state before gunicorn forks workers.
+    if not _tracing_inited:
+        tracing.process_tracing_init()
+        _tracing_inited = True
+
     uvicorn_access = logging.getLogger("uvicorn.access")
     uvicorn_access.handlers.clear()
 

fastdeploy/metrics/trace.py

@@ -289,7 +289,7 @@ def get_otlp_span_exporter(endpoint, headers):
         OTLPSpanExporter as HTTPSpanExporter,
     )
 
-    protocol = os.environ.get(OTEL_EXPORTER_OTLP_TRACES_PROTOCOL, "grpc")
+    protocol = os.environ.get(OTEL_EXPORTER_OTLP_TRACES_PROTOCOL, "http/protobuf")
     supported_protocols = {"grpc", "http/protobuf"}
 
     if protocol not in supported_protocols:

tests/metrics/test_trace.py

@@ -908,6 +908,8 @@ class TestAdditionalCoverage:
 
     def test_get_otlp_span_exporter_grpc(self):
         """Test get_otlp_span_exporter with grpc protocol"""
+        # Set environment variable for grpc protocol
+        os.environ["OTEL_EXPORTER_OTLP_TRACES_PROTOCOL"] = "grpc"
         exporter = trace.get_otlp_span_exporter("http://localhost:4317", None)
         from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
             OTLPSpanExporter as GRPCSpanExporter,