On 02/12/2015 11:04 PM, Quanah Gibson-Mount wrote:
Has anyone done any real testing of Redis as a bayes backend? Talking
with one of our customer, with a trivial <60,000 accounts, they are
seeing:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
22452 redis 20 0 28.7g 28g 740 S 9.6 72.4 1139:38
redis-server
28GB purely for a bayes DB for < 60k accounts is insanity.
2015-02-12 23:19 Axb wrote:
Note: You could be seeing the 28GB of memory usage for 10 users if you
configure it to do so.
My production Redis bayes with about 75k users looks like:
[...]
# Memory
used_memory_human:3.59G
used_memory_peak_human:3.68G
[...]
local_bayes_redis.cf:
bayes_seen_ttl 1d
bayes_token_ttl 5d
Right, keeping _ttl low is the main memory control mechanism.
In particular the bayes_seen_ttl can be kept very low with
no ill-effects. Then use the bayes_token_ttl for adjusting
memory usage - keep it as high as can be comfortably afforded.
The other valuable control is:
bayes_auto_learn_on_error 1
Mark
