Huh, that's bizarre. Has the python version changed? It might be using a different sqlite version again. Can you time how long `.dump` from an `sqlite3` shell takes from each client? In my work we have countless SQLite databases hosted on nfs, with up to a dozen concurrent readers/writers on different machines, and many times the size of yours. We have very few problems due to nfs although the approach is limited in terms of scalability/concurrency. I don't remember any significant slowdowns when we moved from CentOS6 to CentOS7, but that doesn't necessarily translate to RHEL. We're also using nfs4.