Large batches help avoid multiple network round-trips, but I think beyond 500K or 1M it's also a question of how many fields are being indexed and how efficiently we can index them in-parallel.