To tolerate f failures, we need to have f+1 replicas of the data. This implies increased storage, but the f parameter is entirely up to the application.

Record insertion affects all replicas of the data, so latency increases with replication.

The first graph on the performance page (http://hyperdex.org/performance/) shows that a 100% insertion workload will exhibit higher latency (and thus lower throughput for a fixed number of synchronous clients). Write-mostly workloads are slower, but I wouldn't say "extremely slow."