Whitepaper briefly goes over this. The user or application software get to decide their replication number N. The network is self healing, as in if a portion of a file misses a heartbeat(check in) the network will recover it from the other peices.

To affect a file large amount of nodes would have to go down simultaneously. Again the user has control over this N = 1 - Data failure N = 3 - Good for just average cases. Probably still Sybil attacks possible if the have 20-30% of the network. N = 100 - Are you expecting tactical nuclear strikes or acts of God?

Cool thing is if you increase redundancy to a large number Storj starts to act like a CDN or torrent network. But thats outside the scope of your question.