Hello Chris, all,

Wednesday, July 20, 2005, 8:37:20 AM, you wrote:

>> So I think the one-off is more useful for developing a ruleset
>> while the daily full MC is more useful for integration testing.
>> Personally, I participate in both.  With my small corpus it's no
>> biggie either way.   

And with my corpus, with 2 months' emails counting 250k, a single rule
mass-check takes 8 hours. A full system mass-check (that one rule plus
all distribution rules) can easily take 16 hours, especially if that
one rule is subject to performance problems.

CS> one-off corpus checks are a serious backbone to SARE's success. IT
CS> allows us to see which direction the rules should take, what won't
CS> work, and the rules are often coded differently the first 2 "one
CS> off" checks, then changed for a final.   

CS> Its not uncommon to see a rule or ruleset checked 3-5 times before
CS> release. Which often takes less then a day.

Indeed, it's not uncommon for a rule or ruleset to be checked 2-3
times with knowingly excessive regexes, so we can see what actually is
or isn't being matched in various regex hits.  We use this information
to improve the rule, and then remove the excess to the regex for a
final pre-publication run.

CS> Perhaps some thing like the dev "bug squish events" could be used?
CS> Once a week the people who run SARE rule sets check to see the
CS> biggest hitters, and on that day we test those heavy hitters
CS> against a bigger corpus, and look to add to SA. Successful ones
CS> get moved out of SARE and into SA.

That would work for me.  I'm not suggesting that all of these options
be available -- just throwing out ideas:
a) a mailing list for quick turn-around checks. I can probably create
a system here with a special corpus that contains randomly selected 5k
ham and 5k spam from the full 2 month corpus, and turn most of those
mass-checks around in minutes. Chris T and Fred T already have
systems running that turn around mass-checks in minutes. This is a
technique we can probably share (if needed -- ours isn't the only
method).
b) nightly mass-checks run specifically on rules submitted for this
purpose. This would be include rules/70_testing.cf, but not the
entirety of rules/*.cf.  This would enable reasonably quick
mass-checks on all sorts of rules.
c) weekly mass-checks on the full svn trunk rules/*.cf plus, plus
anything submitted to the weekly mass-check (which could be everything
submitted to the nightly mass-check except those that are specifically
pulled). These will run longer, but give better statistics in
comparison between rules.
d) monthly mass-checks of ", with network tests enabled.
e) special network mass-checks of rules that require network access,
as needed.
f) monthly rescoring mass-check for rules that are worth while,
followed by perceptron to rescore rules, and an sa-update to
distribute the rules.

Bob Menschel




Reply via email to