[rules-users] Avoid evaluation of expensive LHS terms

Stumbled over this today in a specific case, but my question is about a general practice (I'm quite new to Drools, 5.2.0.M2 btw). I have a knowledge session with a fair amount of facts (0.5 Mio) and mostly simple rules. One rule has a LHS expression that I know will be expensive to execute, in the example that I encountered this it looks like: rule "Count facts" when all : Number() from accumulate (s : X(), count(s)) blacklisted : Number() from accumulate (s : X(status == X.BLACKLISTED), count(s)) .... then .... end I want to execute this rule only after other rules have already executed (and changed the status of most of the X facts). There's quite a number of X facts (~200'000). Most of them have their status modified during execution of other rules. If I run the rule base without the above rules, performance is perfect, some tens of seconds. If I activate the rule, execution never finishes (I stop it after a few minutes, never saw it finished). I *THINK* (correct me if I'm wrong) what happens is that especially the second expression of the LHS gets executes over and over again whenever one of the facts is modified by one of the other rules. Makes sense, but I nevertheless want to try to avoid it. Until now I tried unsuccessfully: - "protect" the rule with a guard fact as the first expression, which is inserted only late in the processing - move the rule to another ruleflow-group (or agenda-group) and activate it only at the end of the processing - use the rule in a second independent knowledge base and session, and insert the facts once the first has finished I'm perfectly aware that I could solve this specific problem of counting facts with different techniques, but my question is more general (as the subject says). Because I have other cases where (as a last resort) I need to go to a database or online service to get the required information. It seems that others have similar problems (http://drools.46999.n3.nabble.com/Short-circuiting-evaluations-on-LHS-td5...). But in my experience, that technique ("protecting" the expensive pattern in the LHS with a cheap one that evaluates to false until you really need the expensive one) does not work. I get the impression that (contrary to popular experience that a "false" value in an AND sequence terminates the whole sequence) the drools LHS are processed differently. Am I missing something really stupid?! -- CU, Joe