You may know this one from Anki, where it was a plugin.
The goal is to distribute reviews evenly and prevent spikes, to make the workload about the same every day. Especially for higher intervals one or two days off do not interfere with the learning curve, while making the time spent in queue more predictable.
When answering a card, the scheduler checks the number of cards due at the intended date ±X days, choosing the least occupied one for the new interval. X would be to set by the user, where X = 0 means no balancing. There is one X for each phase (Learning/Exponential/Re-learn).