I've heard great things about Pola.rs performance. To get there, they have a lazy evaluation so they can see more of the computation at once, allowing them to implement optimizations similar to those in a SQL engine.
In the early days, even as I appreciated what Pandas could do, I never found its API sane. Pandas has too many special cases and foot-guns. It is a notorious case of poor design.
The problems of the Pandas API were not intrinsic nor unavoidable. They were poor design choices probably caused by short-term thinking or a lack of experience.
On eager vs lazy evaluation -- pytorch defaulting to eager seemed to be part of the reason it was popular. Adding optional lazy evaluation to improve performance later seems to have worked for them.
