On Fri, 14 Feb 2020 10:46:00 -0800 (PST) John Hardin wrote: > On Fri, 14 Feb 2020, RW wrote: >
> > I was looking at it in a bit more detail and it looks like there > > isn't a reader-writer lock, just write locks for the toks and seen > > files. As scans defer their writes through the journal they are > > lockless. > > So, auto-training may be problematic w/r/t locking as well. I presume > that window is coded to be as small as possible. I wasn't referring auto-training, that's done at the end under locking, and picking-up count updates made by other processes since the Bayes classification. The OP has 'bayes_auto_learn 0', so it's not really relevant.