The before_send_span callback is called at _capture_telemetry without wrapping it in capture_internal_exceptions(), so any exception raised by the user's callback propagates all the way to span.end() / the with start_span(): exit, breaking user code. Both before_send (line 873) and before_send_transaction (line 899) are wrapped with with capture_internal_exceptions():; apply the same guard here.

Also found at:

• sentry_sdk/utils.py:2121

When before_send_span returns a dict containing "name" but no "attributes" key, telemetry._attributes is unconditionally cleared and rebuilt from the (missing) key — effectively dropping all span attributes, including sentry-internal ones like sentry.segment.id and sentry.segment.name that are set in _end() just before the span is captured.

In sentry_sdk/consts.py the _experiments typed option declares before_send_span as Optional[Callable[[SpanJSON, Hint], Optional[SpanJSON]]], and get_before_send_span in sentry_sdk/utils.py mirrors this Optional[SpanJSON] return type. However, by design (and per the PR description), before_send_span cannot drop a span. In client.py::_capture_telemetry, when the callback returns None, the code falls through the isinstance(serialized, dict) and 'name' in serialized check, logs "[Tracing] Invalid return value from before_send_span. Keeping original span." at debug level, and re-serializes the original span. This contradicts the type signature, which suggests None is a legitimate return (as it is for before_send, before_send_log, and before_send_metric). Users typing against the public type may reasonably write return None to attempt to drop a span and instead get a silent debug-log warning while the span is still sent. Recommend tightening the type to Callable[[SpanJSON, Hint], SpanJSON] (non-optional return), and updating get_before_send_span accordingly.