CASSANDRA-11465 introduces a new system property cassandra.wait_for_tracing_events_timeout_secs that controls whether or not C* waits for events to be written before responding to client. The current default behavior is to wait up to 1 second and then respond and timeout.
If using probabilistic tracing this can cause queries to be randomly delayed up to 1 second.
Changing the default to -1 (disabled and enabling it explicitly in cql_tracing_test.TestCqlTracing.tracing_unknown_impl_test.
Ideally it would be nice to be able to control this behavior on a per request basis (which would require native protocol changes).