[tomnipotent]

It's min/max per row group, so (potentially) huge chunks of the Parquet file don't need to be read from disk if only a subset qualify.