The conditional logic assigns span.set_attribute when the span is a Span instance, but Span (from sentry_sdk/tracing.py) only has set_data, not set_attribute. Conversely, when the span is a StreamedSpan, it assigns span.set_data, but StreamedSpan (from sentry_sdk/traces.py) only has set_attribute, not set_data. This will cause an AttributeError whenever _record_exception_on_span is called with either span type.

Also found at:

• sentry_sdk/integrations/sqlalchemy.py:102

When an exception occurs during await old_execute_command() at line 150, the code flow exits immediately without calling db_span.__exit__() or cache_span.__exit__(). This causes spans to never finish, leaking them without sending to Sentry. The sync version in _sync_common.py correctly uses try/finally but the async version lacks this protection, leading to orphaned spans and potential scope corruption.

In span streaming mode, when ty is "http" but method is NOT in http_methods_to_capture (e.g., HEAD or OPTIONS requests), the code still creates and starts a span (line 238). In contrast, the legacy non-streaming path correctly returns nullcontext() when these conditions are not met (line 274), avoiding span creation. This causes span streaming mode to trace HTTP methods that users explicitly want excluded, resulting in unwanted spans being sent to Sentry.

When span streaming mode is enabled, the StreamedSpan does not have the http.response.status_code attribute set after the HTTP request completes. The legacy Span.set_http_status() sets SPANDATA.HTTP_STATUS_CODE, but the StreamedSpan path only sets status ('ok'/'error') and 'reason'. This means HTTP status code information (e.g., 200, 404, 500) will not be available in the span attributes for span-streaming users.

Also found at:

• sentry_sdk/integrations/httpx.py:199-201

In sentry_patched_execute_command, if old_execute_command raises an exception, the finally block attempts to call _set_cache_data(cache_span, self, cache_properties, value), but value was never assigned. This causes an UnboundLocalError. While capture_internal_exceptions() suppresses the error (logging it silently), this masks real Redis failures and causes unexpected log noise in the SDK.

When creating NoOpStreamedSpan instances at lines 1302-1304 and 1306-1308 for child spans, the scope=self parameter is omitted, unlike the segment span cases at lines 1265 and 1280 which include it. Without the scope parameter, NoOpStreamedSpan._start() returns early and doesn't set the span on the scope, and _end() doesn't restore the previous span. This breaks the scope's span tracking hierarchy for subsequent span operations.

The constants BAGGAGE_HEADER_NAME and SENTRY_TRACE_HEADER_NAME are defined twice - first in the newly added lines (38-40 area) and again in the existing code (lines 45-46 shown in context). While Python will simply use the last definition, this is clearly unintended code duplication that indicates a merge error or copy-paste mistake. The duplicate definitions waste memory and create confusion for maintainers.