This is the step that never works consistently for me. There is always some amount of random extra latency that makes the this workflow painful. I work with some extremely large data files, so random access to these is the primary issue.

In general, the idea is that it is often better to do compute where the data already is. My experience is that you should also do the programming closer to where the data is as well. This tends to make an iterative development loop tighter.

But this is highly dependent upon what you’re doing.