A data source has to be able to access the metadata to expose its output columns. I haven't verified this, but I don't think the metadata of a data source can change between iterations of a loop. I suspect that when your loop runs the first time, the data source executes the first query and acquires the metadata of the result set. On the next iteration, the second query returns column names that differ from the established metadata. The data source can't resolve the columns of the second result set to the metadata of the expected result set and throws the error.