We largely rely on 8k but sometimes use 32k. Goal is to abstract this away and have things “just work.”