This is so theoretical that not even Alexey (the main CBQ author) claims to understand it. From his source:
David D. Clark, Scott Shenker and Lixia Zhang Supporting Real-Time Applications in an Integrated Services Packet Network: Architecture and Mechanism.
As I understand it, the main idea is to create WFQ flows for each guaranteed service and to allocate the rest of bandwith to dummy flow-0. Flow-0 comprises the predictive services and the best effort traffic; it is handled by a priority scheduler with the highest priority band allocated for predictive services, and the rest --- to the best effort packets.
Note that in CSZ flows are NOT limited to their bandwidth. It is supposed that the flow passed admission control at the edge of the QoS network and it doesn't need further shaping. Any attempt to improve the flow or to shape it to a token bucket at intermediate hops will introduce undesired delays and raise jitter.
At the moment CSZ is the only scheduler that provides true guaranteed service. Another schemes (including CBQ) do not provide guaranteed delay and randomize jitter."
Does not currently seem like a good candidate to use, unless you've read and understand the article mentioned.