I am working on a project where I apply k-means on severals datasets. These datasets may include up to several billion points. I would like to use mini batch k-means to save time. However, the mini batch k-means requires a value for the batch size argument (I am using sklearn). What is the best way to choose a good batch size?