Is there any limitation in establishing BGP peering directly with spine layer instead of the TORs on the edge racks? The topology being considered is a L2 fabric design with MLAG on the spine. the BGP peering VLANs from edge VMs could be extended all the way upto the spine layer and eBGP established between the two spine nodes. This is requirement is arising out of the need to preserve the BGP sessions even if the Edge node VMs are migrated to another rack by vMotion . Hence there wouldn't be any BGP session drop / churn due to this.
What would be the recommendation for a L3 leaf and spine? the VLANs cannot be extended all the way upto spine without VXLAN. So the BGP session should be over a VXLAN tunnel.
Those are logical tunnels and it should work as long as the configurations are done correctly. BTW spine-spine connectivity support confirmation should come from the respective vendor in your case. If they are challenges, you can peer with your firewall devices keeping the fabric design as it is.
You can do that assuming there are redundant links between Spine and Leaf .
What would be the recommendation for a L3 leaf and spine? the VLANs cannot be extended all the way upto spine without VXLAN. So the BGP session should be over a VXLAN tunnel.
Are you referring to underlay VXLAN ?
Great, yes, there are multiple redundant links from leaf.
Yes I was referring to the underlay VXLAN from the leaf to the spine (spine should also support VTEP).
Those are logical tunnels and it should work as long as the configurations are done correctly. BTW spine-spine connectivity support confirmation should come from the respective vendor in your case. If they are challenges, you can peer with your firewall devices keeping the fabric design as it is.