In the same way a slave database can be corrupted by the master when it has a failure. If you're encrypting data on the master and replicating to the slave, improperly encrypted data may be propagated before a fail-over. This is a case where having the master fail completely and go off-line is preferable to a "working failure".
Ironically something not far off from that happened to a contract outfit I was working for. They had a telecoms rack rather than a proper one and they couldn't bolt 9x maxed out DL380G2 servers on it as it started to bend. They decided the best approach was to just pile them up on the rack bed.
This was until the one at the bottom needed a RAM upgrade. Cue 6 people carefully trying to lift up 8x DL380's in one go and sliding various OReilly books underneath. The inevitable happened - a hand gave way resulting in a pile of DL380s spread all over the machine room floor and a broken toe from someone in the hosting facility.