@pieterh that's a very, very interesting diagram from the PDF you've referenced (on page 62).
I wonder, perhaps, if some junior Cisco engineer, saw that two stack ports vs. one stack port, doubles bandwidth?
Un-huh, so only switches 1 and 3 bandwidth is halved, eh? Hmm, walk the path between any two switches, and see what's the bandwidth actually available between them!
Between switches 2 and 4, there's only one link between them! I.e. logically there is now a daisy chain, no longer a ring. Switch 2's other stack port only leads to switch 1, while switch 4's other stack port only leads to switch 3. There's no second path! Indeed, it's now an open ring, again, it's now a daisy chain.
I would really like to see a better explanation how bandwidth between any two switch members is not halved when there's a down stack connection between stack members, anywhere in the stack.
Yes, it's true, for example, stack members 2 and 4 can exchange data with two other stack members while stack members 1 and 3 can only exchange data with one other stack member, but stack members 2 and 4 do NOT have two stack ports bandwidth to each other!
I believe earlier stacks operated the same way.
Also, interestingly, on the prior page of the referenced PDF, shows how stack member can interconnected. It's sort of implied this is something new with the C9350, but it's not.
As a side note, when Cisco introduced StackWise, with the 3750 series, it was advertised as 32G. Well, as best I could tell, each stack port was actually 8G, duplex. Cisco didn't actually provide much detail on how the stack ring operated.
However, with the introduction of the 3750E series, and its StackWise+, they noted it was now 64G (I believe each stack port was now 16G, duplex). They also noted a couple of "improvements", besides bandwidth, with StackWise+. Firstly, they now described, at least for unicast traffic, it was only placed on the ring, when it needed to be placed on the ring. (I.e. this implying the original StackWise, flooded all traffic on the stack ring, even if between two switch ports on the same physical switch member.) Secondly, the destination switch removed unicast traffic, from the ring, "spatial reuse". (I.e. implying the original StackWise source switch removed traffic, that it placed on the ring, after it fully transited all the other switches!)
Until encountering the StackWise+ enhancements, it was unclear how yucky original StackWise was, at least in stack ring bandwidth management. It's certainly possible, there may be improvements in StackWise, for 9Ks, possibly specifically for C9350s, but it's unclear how page 62 is actually an improvement. (Although, also on page 62, mention is made of packets being fragmented and sent across both ring paths, when available. Yet the C9350 isn't impacted the same way with the ring being broken!?)
BTW, I only mention the 3750 and 3750E, because, Cisco, like other sellers, often touts "features" that aren't much of a feature, at least when it comes to performance. Back in the 3750 day, it only had FE edge ports (the later 3750G had gig), and Gig was still rather novel, so a 32G ring, was wow! (But was it really? Laugh, another StackWise sales "feature" was when a max stack member stack, sum of all its stack ports' bandwidth, was touted.)
Caveat emptor.
