On Tue, 7 Nov 2017 00:59:12 +0300 David Gessel wrote: > FreeBSD is currently installing TxRep.pm rev 1651114 from Jan 12 > 15:17:46 2015 (it is the only revision that has only whitespace > differences, all leading padding, there are code differences between > installed and 1650327 (previous) and 1678017 (next). The most recent > is 1720440 from Dec 16 20:23:02 2015. > ... > > I'll see if anyone can make more sense of this than I can. If not, > I'll try each of the TxRep.pm variants from the one that is fast but > throws errors to the current one and see if that provides any insight > into the problematic code change.
> It should significantly speed up sa-learn if it were threaded, no? You can do that using multiple spamc processes with spamd, but it doesn't use any less resources, and the default back-end has a reader-writer lock. Training with sa-learn is usually not an issue compared with high-volume autotraining (which is naturally parallel). I'm using SA from the FreeBSD port on a low-end i5 from a few years back with the default Berkeley db on a fairly slow disk and I get 20 emails/s through sa-learn (without learning to journal). However, I don't use TxRep. Have you tried commenting-out the TxRep loadplugin line, to see what happens? In your original post it wasn't simply slow, if it had continued at ~1600/hour, it would have completed in a few hours, but it ground to a halt and apparently the Bayes database got corrupted. If TxRep is the source of the problem then disabling it on your server while you debug it separately would be sensible IMO.
