Glad you liked to document.

I would say it heavily depends on your requirements and the overall architecture meaning e.g. how many spine and leaf switches do you have and do you plan to implement in your environment, where do you want to manage L2 vs L3 traffic.
If you connect them you’ll have a MLAG type of network and if that is what you want you can definitely go for it. Just be aware of configuration required to connect the leaf to the spine so you don’t end up in a troublesome situation.

//Magnus

In a collapsed core environment, should the Leaf switches (those directly connected to each Nutanix node) be connected to each other?

Consider the following scenario:
10+ nodes (but less than 48)
Each w 2 x 10Gb SFP+ uplinks (Call ’em: Uplink A & B)
Two N9K-C9372PX-E w NX-OS 7.0+ (Call ’em Leaf Switch A and B)
Hypervisor is ESXi and is using a dedicated VLAN for all CVMs w a VDS.

We’ve experience a situation where let’s say Node A’s CVM traffic is placed on Uplink A, while Node B’s CVM’s traffic is placed on Uplink B.
In this scenario in order for CVM A and CVM B to communicate, they have to CROSS the spines to get to each other. (Leaf Switch A -> Spine Switch A -> Leaf Switch B)
(See “Leaf-Spine Network” figure your document)

Consider the following event:
When Leaf Switch A is rebooted-ed (on purpose) or it’s firmware updated, the two nodes have a momentary decision to make to fail VM traffice from Uplink A to Uplink B.
In our experience this failover process takes long enough for CVMs to notice (CLuster was fine, but had some warnings in “/home/nutanix/data/logs/sysstats/ping*” from CVMs due to reachability)

Would one solution be to administratively down the Uplinks on Leaf Switch A first, and let all uplinks fail to Leaf Switch B “nicely” before rebooting Leaf Switch A?

Would you agree that it would be better to create a (40Gbps) link between Leaf Switch A and Leaf Switch B (limit traffic for ESXi/CVM’s VLAN of course?)

Hi Craig, we can take this discussion offline. It’s basically about knowing what features are in use across the customer network but good point.

thanks

good stuff – glad you found it useful.

//magnus