One limitation I find annoying is that you can't easily strip your transformation from the data and apply it to similar but different data. It's all together "a project."

So incrementally transforming data works well, applying that transformation elsewhere not really. That bothered me more than the size limitation, which I think is not limiting most use cases with real “messy” data anyway. Maybe processing large volume event / log data would still need something else.

I used it to look at a json export from a SaaS tool, and to convert it to table structure. Cleaning field contents, which where following certain conventions, but which evolved over time, things like that. For such use-cases it's powerful.