...
You could spend a lot of time trying to train your secondary box(es), but that's a lot of unnecessary work. Instead, use the Bayes knowledge that your primary box has generated by copying over all of the bayes_* files from the primary to the secondary boxes, making sure all the rights are correct, and restarting. You could do this on a regular basis manually, or you could get clever and figure out a way to do it automatically using clustering, FTP, or such. Regardless of how you do it, it's a must if you truly want fault tolerance in your SpamAssassin/Bayes email filtering system.
SpamAssassin 3.0.0 includes an SQL backend for Bayes – this would allow multiple servers to share a single Bayes DB efficiently. If you are running this configuration with SQL Bayes, please share your experiences here.