We currently try to decode everything. Of course we silently fail in case the file is a binary, pdf, or whatever can not be decoded to UTF-8.
Instead of silently failing, we could return an undecoded version of the file.

idea

https://duckdb.org/ is a hot new tool. It could be a nice way to load data for people. We should show some ways people could use it with Hamilton. IIRC duckdb does predicate push down when querying files like parquet, which can be faster than loading them via pandas and then doing some filters.

Rough sketch:

1. data_loaders module that connects with duckdb
2. feature_logic module(s)