The role of BGP inside an AS is to disseminate the routes learned from external peers to all routers of the AS. A straightforward, but not scalable, solution, is to resort to a full-mesh of iBGP sessions between the routers of the domain. Achieving scalability in the number of iBGP sessions is possible by using Route Reflectors (RR). Relying on a sparse iBGP graph using RRs however has a negative impact on routers’ ability to quickly switch to an alternate route in case of a failure. This stems from the fact that routers do not often know routes towards distinct next-hops, for any given prefix. In this paper, we propose a solution to build sparse iBGP topologies, where each BGP router learns two routes with distinct next-hops (NH) for each prefix. We qualify such iBGP topologies as NH-diverse. We propose to leverage the “best-external” option available on routers. By activating this option, and adding a limited number of iBGP sessions to the existing iBGP topology, we obtain NH-diverse iBGP topologies that scale, both in number of sessions and routing table sizes. We show that NH diversity enables to achieve sub- second switch-over time upon the failure of an ASBR or interdomain link. The scalability of our approach is confirmed by an evaluation on a research and a Service Provider network.