bayes_toks corrupted

[Jim Flowers]

I am using mailscanner 4.26.8-1 on a fairly high volume mail relay (~6000
messages per day) and MailWatch on a FreeBSD 4.9 machine.  Everything
seems to be working OK but after a couple of days, bayes stops working and
sa-learn cannot be used to correct it.

Normally, bayes_toks is a db file that `file` returns:

     bayes_toks:    Berkeley DB 1.85 (Hash, version 2, native byte-order)

When it stops working, `file` returns:

     bayes_toks     data

and I have to start over again with a good bayes_toks file.

Has anyone else run into this and solved it?  Would it help to use DB3,
DB4 or DB4.1?

Thanks.