Schemas are overrated. Often the source-system can't be trusted so you need to check everything anyway or you'll have random strings in your data. Immature languages/libraries often do dumb stuff like throwing away the timezone before adjusting it to UTC. They might not support certain parquet types (e.g. an interval).

Like I've recently found it much easier to deal with schema evolution in pyspark with a lot of historical CSVs than historical parquets. This is essentially a pyspark problem, but if everything works worse with your data format then maybe it's the format that's the problem. CSV parsing is always and everywhere easy, easier than the problems parquets often throw up.

The only time I'd recommend parquet is if you're setting up a pipeline with file transfer and you control both ends... but that's the easiest possible situation to be in; if your solution only works when it's a very easy problem then it's not a good solution.