The Definitive Guide to Data Quality
For your description, I think the performance depends on your work flow, your data and the transformations.Could you give us more information about your ETL so that we can design a proper job for you.
Consider a use case where, to loading my fact tables i have to join them with 5 dimension tables (like geography, customer, product, etc.) and pull their ID's into my fact table. Now since the data is huge (~ 50 million records per day) so doing it from tMap (that too storing the data on disk to avoid Java Heap space issues) will be slower than writing an equivalent SQL for this use case and triggering it from tMSSqlRow.