Description
Spark core computes callsites by climbing up the stack until we reach the stack frame at the boundary of user code and spark code. The way we compute whether a given frame is internal (Spark) or user code does not work correctly with the new dataframe API.
Once the scope work goes in, we'll have a nicer way to express units of operator scope, but until then there is a simple fix where we just make sure the SQL internal classes are also skipped as we climb up the stack.