https://issues.apache.org/SpamAssassin/show_bug.cgi?id=6942
--- Comment #15 from AXB <axb.li...@gmail.com> --- (In reply to Mark Martinec from comment #14) > > > > 12240246 non-token data: nspam > > > > 5076877 non-token data: nham > > > > used_memory_human:3.10G > > > Suggesting: bayes_auto_learn_on_error 1 > > > > why? > > From the stats it seems to me like a large number of tokens, > and 3 GB of resident storage is on a high side and probably > growing still at the same rate (until expiration kicks in). This is two weeks' worth of tokens, running on a dedicated redis box with 32 gb of ram. Since last Redis server upgrade memory usage has decreased quite a bit. I was peaking 4.7 GB of Redis usage. DB size is pretty stable so it seems expiration has been reliable. 99% of spam is fed via traps - NOT production flow. ham is production ham, in autolearn mode and there's no false learning. (autolearn_force does wonders with ham and spam) > The bayes_auto_learn_on_error can reduce the growth rate > substantially, without sacrificing much on the quality of > results. Some studies even indicated that a learn_on_error > strategy increased the classification quality (but I won't > speculate on that here). I'm not worried about size - and speed is so fast that it can only be beat by turning off bayes completely. As to classification quality, I see no errors, in neither way. Don't quite see how this setting could make it even better, but open to education -- You are receiving this mail because: You are the assignee for the bug.