Each json file is of a single object and has the potential to have variance in the schema.

How much variance are we talking?  JSON->Parquet is going to do well with 100s of different columns, but at 10,000s many things will probably start breaking.