This is completely from a Python neophyte's perspective but, based on discussions with developer regarding other IDEs, functions are and libraries are great! They provide functionality in reference call reducing the amount of time required to develop the same functionality manually. There is cost for that convenience though; you have memory overhead required for preloading libraries and other add-ons and then you have reference lag (looking up and loading the function) which you don't have with task specific code written out long hand (so to speak). With today's processing speeds and I/O capacity, many will poopoo this but in my discussions with long term coders in the MS Visual studio field, the dislike of bloated libraries and dlls, and the overhead and performance hits endemic with (dot)NET libraries are just something you have to deal with, otherwise, you have to roll your own leaner meaner utilities.

I agree that you can't test with a few records and make a broad generalization like you have, even a warm breath from the fan on a resistor could be responsible for your perceived performance inequities. Run the same test against a half a million records, then run it again after resequencing you process executions to give each process the opportunity to be first/second/third, then come back with your results.

Personally, my bias (neophyte-bias) tells me you may be right but my curiosity thinks a better test is in order.