I'd say it's like anything else. It depends how you are bound. If you can run them on the same machine, go ahead..but if you become resource bound, you split them up, then you shard...common scaling approach.

As for why not use native structures? You need to build persistence, transaction support, pub/sub api, thread awarness. Also, if you happen to use a generic data structure, like Java's HashSet, you probably won't get even close to the same level of performance.