Leaf/Spine State Management Is Critical
If processing capacity is fixed, then what it is being processed also has to be fixed, otherwise processing time will increase.
Figure 1 depicts the three olive martini model of network dynamics, with the network component exploded to Network Design, Control Plane, and Physical Network.
Figure 1. Network components in he three olive martini network dynamics model
While the control plane is ultimately responsible for path selection, it is important to note that a) in order to do that well & fast, it needs to understand the state of the physical network and b) how the control plane operates is impacted by design: protocol options, network architecture, equipment design, etc.
One of the critical issues in networking is how much state is in the control plane, how fast it can be processed, and if there is enough memory to hold the state. In a network based on fixed configuration platforms, where the processor and memory cannot be upgraded, state management strategies are likely to become important. The conventional wisdom on network design is that the core of a network should be about high-speed forwarding. This is mostly driven by a desire to have a stable core, hence one of the drivers for segment routing. It is probable, IMO, that leaf/spine architectures using fixed configuration processing and memory would also drive this trend.
Figure 2. Function time: resource capacity & information volume/velocity
There are a few different ways that network designers could respond to this issue:
Request equipment suppliers have field replaceable processors/memory.
Select control plane protocols that have the least state.
Aggregate / summarize state to reduce volume.
Option 1 would drive up costs, and would only be applicable to those segments where equipment is not already a rip and replace commodity. Design can also hit problems when not all nodes have the same processing / memory capacity and decisions are made based on the highest capacity nodes in the network. The other nodes can become overwhelmed. A variant of option 1 would be chassis-based architectures.
Option 2 obviously brings in the BGP vs link-state protocol debate. BGP for less state and less topology information, but greater policy expression. OSPF/IS-IS for more topology/TE information, but with less policy expression and more state. The choices made here also have Operations impact. Segment routing comes into play as well, however, there may be significant demands on the edge of the network depending on what policy-based routing is deemed necessary (what granularity of TE).
Option 3 is considered a best practice in wide-area networks, a distribution layer, but potentially comes at the cost of path selection optimality in a connectivity rich environment.
Options 1,2, and 3 are likely candidates for consideration in fixed-form factor based networks.
What else would you consider?