https://github.com/apache/datafusion/pull/20822 allows us to push struct field filters down to the parquet source. It would be good to have a series of benchmarks so we can measure the impact