fix(tests): prevent background uploads from racing VCR replay

py/src/braintrust/test_context.py

@@ -308,17 +308,12 @@ def blocking_work():
     test_logger.flush()
     logs = with_memory_logger.pop()
 
-    # SURPRISING: Even to_thread() loses logger context (logger is a ContextVar too!)
-    # Only parent span is logged
-    # However, to_thread() DOES preserve span parent context
-    assert len(logs) >= 1
-
-    # If both spans logged (logger context preserved), verify parent chain
-    if len(logs) == 2:
-        parent_log = next(l for l in logs if l["span_attributes"]["name"] == "parent")
-        worker_log = next(l for l in logs if l["span_attributes"]["name"] == "to_thread_worker")
-        assert worker_log["root_span_id"] == parent_log["root_span_id"]
-        assert parent_log["span_id"] in worker_log.get("span_parents", [])
+    # The memory logger must capture worker-thread exports as well as the parent.
+    assert len(logs) == 2
+    parent_log = next(l for l in logs if l["span_attributes"]["name"] == "parent")
+    worker_log = next(l for l in logs if l["span_attributes"]["name"] == "to_thread_worker")
+    assert worker_log["root_span_id"] == parent_log["root_span_id"]
+    assert parent_log["span_id"] in worker_log.get("span_parents", [])
 
 
 # ============================================================================

py/src/braintrust/test_helpers.py

@@ -60,9 +60,12 @@ def with_simulate_login():
 
 
 @pytest.fixture
-def with_memory_logger():
+def with_memory_logger(monkeypatch):
     logger._state.reset_parent_state()
     with logger._internal_with_memory_background_logger() as bgl:
+        # The override is thread-local. Worker threads need the same fallback
+        # so they cannot start HTTP uploads that outlive the test and race VCR.
+        monkeypatch.setattr(logger._state, "_global_bg_logger", LazyValue(lambda: bgl, use_mutex=False))
         yield bgl
     # Clean up global state to prevent test contamination
     logger._state.reset_parent_state()

py/src/braintrust/test_logger.py

@@ -2484,7 +2484,7 @@ def reset_id_generator_state(monkeypatch):
     monkeypatch.delenv("BRAINTRUST_LEGACY_IDS", raising=False)
 
 
-def test_span_with_otel_ids_export_import(reset_id_generator_state, monkeypatch):
+def test_span_with_otel_ids_export_import(with_memory_logger, reset_id_generator_state, monkeypatch):
     """Test that actual Span objects with OTEL IDs can export and be used as parent context."""
     init_test_logger(__name__)
     monkeypatch.setenv("BRAINTRUST_OTEL_COMPAT", "true")
@@ -2502,7 +2502,7 @@ def test_span_with_otel_ids_export_import(reset_id_generator_state, monkeypatch)
         assert imported.root_span_id == span.root_span_id
 
 
-def test_span_with_uuid_ids_share_root_span_id(reset_id_generator_state, monkeypatch):
+def test_span_with_uuid_ids_share_root_span_id(with_memory_logger, reset_id_generator_state, monkeypatch):
     """Test that legacy UUID generators share span_id as root_span_id for backwards compatibility."""
     monkeypatch.setenv("BRAINTRUST_LEGACY_IDS", "true")
     init_test_logger(__name__)
@@ -2908,6 +2908,9 @@ def setUp(self):
         metadata_patch.start()
         self.addCleanup(metadata_patch.stop)
         self.state = logger.login_to_state(api_key=logger.TEST_API_KEY)
+        # These tests use a separate state, so the global memory override does
+        # not apply. Keep their span exports local too.
+        self.state._global_bg_logger = LazyValue(logger._MemoryBackgroundLogger, use_mutex=True)
         self.state.org_name = "test-org"
         api_client = MagicMock()
         client_patch = patch.object(self.state, "api_client", return_value=api_client)

py/src/braintrust/test_span_components.py

@@ -10,6 +10,7 @@
 from braintrust.id_gen import OTELIDGenerator
 from braintrust.span_identifier_v3 import SpanComponentsV3, SpanObjectTypeV3
 from braintrust.span_identifier_v4 import SpanComponentsV4
+from braintrust.test_helpers import with_memory_logger  # noqa: F401 # type: ignore[reportUnusedImport]
 
 
 class TestSpanComponentsV3:
@@ -346,7 +347,7 @@ class TestExportFormatSelection:
     BRAINTRUST_LEGACY_IDS) serialize as V3.
     """
 
-    def test_export_format_based_on_env_variable(self):
+    def test_export_format_based_on_env_variable(self, with_memory_logger):
         """Test that export format follows the legacy-UUID opt-out flag."""
         from braintrust.test_helpers import init_test_logger, preserve_env_vars
 
@@ -374,7 +375,7 @@ def test_export_format_based_on_env_variable(self):
             assert SpanComponentsV4.from_str(export_v3_mode) is not None
             assert SpanComponentsV4.from_str(export_v4_mode) is not None
 
-    def test_export_uses_v4_by_default(self):
+    def test_export_uses_v4_by_default(self, with_memory_logger):
         """Test that export uses V4 format by default (no env vars set)."""
         from braintrust.test_helpers import init_test_logger, preserve_env_vars