[j-nsp] Design assistance with routing instances

[Paul S.]

Hi folks,

Dealing with some rather interesting design issues lately.

The main issue is that we have to partition the global table to provide 
"selective access"  (mainly due to massive cost disparity based on where 
traffic goes, this is in Asia) to customers as they want. Prefer to do 
it within a single MX router with routing instances.

Usual criterion is like Global / Global + Local /  Local Only and so on.

The issue here is that if I replicate the global table to a 
routing-instance FIB, I basically end up with 1.3m routes there and  
this doesn't seem very future proof.

With that in mind, I guess I'd love some input on:

1. Should I keep all upstream links on the default routing instance and 
only leak to the target vrfs? Or is putting upstreams that only serve 
specific purposes inside specific routing instances a better solution idea?

2. On the topic of route leaking, are rib groups preferred, or real bgp 
sessions over the lt interface? I'd like the ability to announce things 
to "all upstreams" from the router itself.

We plan to develop a backbone network with MPLS-TE for L{2|3}VPN 
services at a later date too, so the design shouldn't cause any 
interoperability issues.

Any input on architecting this is welcome, thank you for reading!