This fix seems to be only partial: consider for example the case when one learner is much cheaper than all the others. Then as soon as one point arrives, the balancing learner will request all points from another random learner in the queue.

If I understand correctly, an empty learner has both an infinite loss and an infinite loss improvement, however some non-empty learners also have the same loss and loss improvement. In that case the problem would arise as well.

Therefore it seems to suggest that a better fix would be to sort on(-loss, n_points) (or loss improvement, depending on the balancing learner setting).

@akhmerov, good suggestion!

I've done this in6e172ef. This is what you meant right?

I'll also do it in_ask_and_tell_based_on_loss.

My problem right now is that the logic is rather nontrivial, and it's hard to follow especially due to very similar variable names. Comments might also help.

I found another issue that I fixed.

I've simplified it now and I think the code is understandable, do you agree?