After some great offline discussions with various folks involved in the specification of HTTP/2 priority, I wrote an email to the list expressing my concern about the design as written.

In response, Martin Thomson agreed with my concerns and proposed a small tweak to the spec that allows for the allocation of dummy nodes, which enables rough approximation of a priority hierarchy.

The tweak makes it possible to send a PRIORITY frame to a stream in any state. Given two streams {A, B} at a higher priority than streams {C, D}, this allows construction of a tree as follows:

A{w=256} -> 0
B{w=256} -> 0
Dummy{w=1} -> 0
C{w=256} -> Dummy
D{w=256} -> Dummy

We can probably assume that a well-behaved server would not spend any resources on the Dummy subtree -- since it has weight=1 -- until A and B are completed, which places C and D at a lower priority.

This protocol tweak appears to be headed for acceptance. Here are the relevant meeting minutes.

Offline, several people expressed general nervousness about HTTP/2 priority, as there is no production experience with a tree-based priority model. I was told that, should HTTP/2 priority not work out in practice, it will be deprecated and some kind of alternate protocol or extension will be proposed instead.

Ultimately, I'm thrilled that, with Martin's tweak, it's plausible that HTTP/2 priority can solve most use cases. I do wish I'd gotten involved in this topic nine months ago, as I would have had a realistic shot at getting Microsoft's proposal adopted. At this point, the standard will ship with only minor changes.