With so many chunks, there’s a chance that on restarting a client (which has a new address and forms a new network view), that it differs in the view of the network, and die to some churn, it doesn’t see the data there.
This should be improved with GetRange work, which will (be in main so soon hopefully), and will sample a larger network space for replication and storage (so more like “minimum 5 peers”, but actually we need to cover a % space of the network, and will ask that larger space too)
There’s quite a lot of reasons we could not find a chunk at a specific time (Packets can be dropped all the …). If it’s been uploaded, the chances are we’ll be able to get it after some replication, or retries. If there’s been heavy churn, it may take a bit.
(the unwanted 
i know! but I don’t see this as a fundamental network issue so much as tweaking of tolerances both network and client side; we’ll firing up more tough churn tests with home nodes too and sampling things to check and assert tolerances over such things. So more rigor here is inbound)
Not necessarily see this thread for more