Polars author here. I have run the TPC-H benchmark against polars and pandas 2.0 backed by arrow types.

https://github.com/pola-rs/tpch/pull/36

Pandas having arrow as backend is great and will make interop with the arrow community (and polars) much better.

However, if you need performance, polars remains orders of magnitudes faster on whole queries, changing to the arrow memory format does not change that.