The training phase is basically pre-compute all the posterior probabilities plus P(c1), P(c2) ..., so in the testing phrase you can simply look up the probabilities instead of computing them again and again for each record.