For tabular/time-series greater than 100k rows I personally feel like parquet cannot be beat. It's self-describing, strongly-typed, relatively compact, supports

For tabular/time-series greater than 100k rows I personally feel like parquet cannot be beat. It's self-describing, strongly-typed, relatively compact, supports a bunch of io/decode skipping, and is quite fast. Orc also looks good, but isn't well supported. I think parquet is optimal for now for most analytical use-cases that don't require human readability.