On 02/13/2015 01:57 PM, Jan Hejl wrote:
Dne 13.2.2015 v 13:31 Henrik Krohns napsal(a):
On Fri, Feb 13, 2015 at 11:12:38AM +0100, Jan Hejl wrote:
Dne 13.2.2015 v 09:47 Axb napsal(a):
On 02/13/2015 08:51 AM, Jan Hejl wrote:
Whups. 62M :-D Sorry, my mistake. But still ...
what counts is:
used_memory_human:49.70M
seems amazingly little data.
Why not?
What counts is tokens (and ttl to purge them). It makes absolutely zero
common sense that 10M messages could be processed having <1M tokens in
the
DB.
I process only few hundred messages a day:
$ redis-cli info | egrep '(:keys|memory)'
used_memory:26683624
used_memory_human:25.45M
used_memory_rss:31256576
used_memory_peak:27370904
used_memory_peak_human:26.10M
used_memory_lua:109568
db0:keys=266576,expires=266572,avg_ttl=2305894191
Well 10M is spread over 5 same SA machines, each runs redis instance -
2M per machine - about 25% SA results are cached on client side (these
are not sent to spamd) - thus it gives 1,5M per machine / SA node (32
cores, 128G memory, LOAD between 8-12, few non-default SA plugins). Each
node can do 5M per day without being overloaded.
Redis output:
used_memory:52039312
used_memory_human:49.63M
used_memory_rss:59494400
used_memory_peak:52873144
used_memory_peak_human:50.42M
used_memory_lua:159744
db0:keys=375966,expires=375962,avg_ttl=3894540722
Bayes settings:
bayes_token_ttl 60d
bayes_seen_ttl 14d
bayes_auto_expire 1
bayes_auto_learn 0
I don't think that it's neccessary to use a huge token base - sometimes
it's about quality, not quantity.
if you use a central Redis for a bunch of spamd boxes and only use
autolearn, quantity is hardly controllable.
