When training the Bayesian filters for SpamAssassin, be very careful about using sa-learn on your active IMAP inbox. Most IMAP clients don’t actually delete messages until they synchronize the mbox. So, after deleting around a hundred Spams, I ran sa-learn over my inbox and SpamAssassin happily learned all of the Spams I had marked deleted as Ham (non-Spam).
Since it was a few hours before I realized what had happened when the incoming Spam getting by SpamAssassin increased tenfold, it was too late to run sa-learn over the old Spams because they had already been truly deleted.
I had to wipe my Bayesian database and start over. No big deal; just inconvenient.