Add propagate-only distributed tracing

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Copilot-Session: b872ac61-5bac-40e3-b9c7-daf295725fde

Author: andystaples

CHANGELOG.md

@@ -9,6 +9,10 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ADDED
 
+- Added `TaskHubGrpcWorker(..., emit_trace_spans=False)` for hosts that own
+Durable Task lifecycle spans. In this propagate-only mode, incoming W3C trace
+context remains active while orchestrator, activity, and entity user code runs,
+without emitting duplicate worker lifecycle spans.
 - Added asynchronous `AsyncScheduledTaskClient` / `AsyncScheduleClient` and
 `AsyncExportHistoryClient` / `AsyncExportHistoryJobClient` APIs. Applications
 using `AsyncTaskHubGrpcClient` can now manage scheduled tasks and history

azure-functions-durable/CHANGELOG.md

@@ -9,6 +9,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ADDED
 
+- Distributed tracing now correlates OpenTelemetry spans created by orchestrator
+user code with the Durable Functions host trace while avoiding duplicate
+orchestration, activity, and entity lifecycle spans from the Python worker.
 - Added `SyncDurableFunctionsClient`. `DFApp.durable_client_input()` now injects
 the synchronous client into synchronous functions and the asynchronous client
 into coroutine functions. Both clients support scheduled-task and history-export
@@ -178,8 +181,3 @@ code:
   `show_history_output` flags for signature compatibility but ignore them, so
   the returned status has no `historyEvents`. Use
   `get_orchestration_history(...)` to retrieve history.
-- Distributed tracing is not yet wired up. The Durable Functions host delivers
-  the parent trace context and emits the orchestration/activity spans itself,
-  so orchestrator user-code spans in the Python worker are not yet correlated
-  to it, and durabletask's own span emission is intentionally left disabled to
-  avoid duplicating the host's spans.

azure-functions-durable/azure/durable_functions/worker.py

@@ -40,7 +40,10 @@ def __init__(self) -> None:
         # The Functions converter routes payload serialization through the
         # azure-functions codec (df_dumps/df_loads) so user types round-trip in
         # the wire format the Durable Functions host extension expects.
-        super().__init__(data_converter=DEFAULT_FUNCTIONS_DATA_CONVERTER)
+        super().__init__(
+            data_converter=DEFAULT_FUNCTIONS_DATA_CONVERTER,
+            emit_trace_spans=False,
+        )
 
     def add_named_orchestrator(self, name: str, func: task.Orchestrator[Any, Any]) -> None:
         self._registry.add_named_orchestrator(name, func)

azure-functions-durable/pyproject.toml

@@ -27,7 +27,7 @@ requires-python = ">=3.13"
 license = {file = "LICENSE"}
 readme = "README.md"
 dependencies = [
-    "durabletask>=1.8.0",
+    "durabletask[opentelemetry]>=1.9.0",
     "azure-identity>=1.19.0",
     "azure-functions>=2.3.0b2"
 ]

durabletask/internal/tracing.py

@@ -18,6 +18,7 @@
 import random
 import time
 from contextlib import contextmanager
+from contextvars import ContextVar
 from datetime import datetime
 from typing import TYPE_CHECKING, Any
 
@@ -85,6 +86,10 @@ class _StatusCodeStub:
 
 # The instrumentation scope name used when creating spans.
 _TRACER_NAME = "durabletask"
+_SPAN_EMISSION_SUPPRESSED: ContextVar[bool] = ContextVar(
+    "durabletask_span_emission_suppressed",
+    default=False,
+)
 
 
 # ---------------------------------------------------------------------------
@@ -226,6 +231,35 @@ def extract_trace_context(proto_ctx: pb.TraceContext | None) -> Any | None:
     return ctx
 
 
+@contextmanager
+def use_trace_context(trace_context: pb.TraceContext | None):
+    """Make a protobuf trace context current without creating a span."""
+    parent_ctx = extract_trace_context(trace_context)
+    if parent_ctx is None:
+        yield
+        return
+
+    token = otel_context.attach(parent_ctx)
+    try:
+        yield
+    finally:
+        otel_context.detach(token)
+
+
+@contextmanager
+def suppress_span_emission(suppress: bool = True):
+    """Temporarily suppress spans emitted by this module."""
+    if not suppress:
+        yield
+        return
+
+    token = _SPAN_EMISSION_SUPPRESSED.set(True)
+    try:
+        yield
+    finally:
+        _SPAN_EMISSION_SUPPRESSED.reset(token)
+
+
 @contextmanager
 def start_span(
     name: str,
@@ -254,6 +288,11 @@ def start_span(
         yield None
         return
 
+    if _SPAN_EMISSION_SUPPRESSED.get():
+        with use_trace_context(trace_context):
+            yield None
+        return
+
     parent_ctx = extract_trace_context(trace_context)
 
     if kind is None:
@@ -313,7 +352,7 @@ def emit_orchestration_span(
     Falls back to ``tracer.start_span()`` (which generates its own
     span ID) when *orchestration_trace_context* is ``None``.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return
 
     span_name = create_span_name(TASK_TYPE_ORCHESTRATION, name, version)
@@ -529,7 +568,7 @@ def generate_client_trace_context(
     returned the caller should fall back to the orchestration's own
     trace context as the parent for downstream spans.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return None
     if parent_trace_context is None:
         return None
@@ -588,7 +627,7 @@ def emit_client_span(
     SERVER span remains connected to the orchestration span via the
     fallback parenting established in :func:`generate_client_trace_context`.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return
 
     # SDK-internal imports — see _is_deferred_span_capable() for details.
@@ -698,7 +737,7 @@ def emit_timer_span(
     When *parent_trace_context* is provided the span is created as a
     child of that context; otherwise it inherits the ambient context.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return
 
     span_name = create_timer_span_name(orchestration_name)
@@ -741,7 +780,7 @@ def emit_event_raised_span(
     When *parent_trace_context* is provided the span is created as a
     child of that context; otherwise it inherits the ambient context.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return
 
     span_name = create_span_name(SPAN_TYPE_ORCHESTRATION_EVENT, event_name)
@@ -787,7 +826,7 @@ def start_create_orchestration_span(
     Yields the span; caller should capture the trace context after entering
     the span context so it can be injected into the gRPC request.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         yield None
         return
 
@@ -815,7 +854,7 @@ def start_raise_event_span(
     target_instance_id: str,
 ):
     """Context manager for a Producer span when raising an event from the client."""
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         yield None
         return
 
@@ -845,7 +884,7 @@ def reconstruct_trace_context(
     the span ID with *span_id*.  This is used to reuse a pre-determined
     orchestration span ID across replays.
     """
-    if not _OTEL_AVAILABLE:
+    if not _OTEL_AVAILABLE or _SPAN_EMISSION_SUPPRESSED.get():
         return None
 
     parsed = _parse_traceparent(parent_trace_context.traceParent)
@@ -869,7 +908,7 @@ def build_orchestration_trace_context(
     replays so that all dispatches produce a consistent orchestration
     SERVER span.
     """
-    if start_time_ns is None:
+    if start_time_ns is None or _SPAN_EMISSION_SUPPRESSED.get():
         return None
 
     ctx = pb.OrchestrationTraceContext()