Typically PMs are installed in a redundant pair, each PM runs hot and is capable of assuming the full processing load should a failure occur.
All single point failures and some multi-point failures are handled by this architecture.
Most failure modes are transparent to the client application, however a client application must be preapred to switch to another PM should the PM to which it is connected to fail.