Yeah, clients always expect CSV (or sometimes XLSX), but if I tell them that I'll send parquet data, they will ask if I'm having a stroke or something because they don't know what is parquet and how could they use it.

CSV is just too simple and "user-friendly".

Oh, yes. And even if you can convince the client that they're wrong, and you're right - with substantial client datasets, there's always a load of "data not as previously represented" records. Resolving what is going on with those tends to be vastly easier when you can say "look at record 1,234,567" and they can easily do that in their favorite & familiar software.

Moreover, I myself like being able to open broken csv files in a text editor, to find nulls and other problematic junk.