1:36 a.m.
I am making extensive use of the event processing features of the Drools
rule engine. Upgrading from Drools 5.2.0.Final to Drools 5.3.0.Final broke
47 of my unit tests and also broke my functional tests. There seem to be
multiple changes in Drools 5.3.0 that cause incorrect behavior and/or break
backward compatibility. Here are the results of my investigation so far.
Issue F1:
While tracking down the failures of my unit tests, I experimented with the
Broker example, and while probably not finding the root cause of the broken
unit tests, I nonetheless came across what clearly seems to be incorrect
behavior in the DefaultTimerJobInstance.call() method. The bug only reveals
itself after all input has been processed, so it is masked by the fact that
running the Broker demo through the entire sequence in stocktickstream.dat
(1100 lines) takes a long time. In order to reveal the problem more
quickly, run the demo with only the first 14 lines in stocktickstream.dat.
Do so for both the 5.2.0 Broker demo (against Drools 5.2.0) and the 5.3.0
Broker demo (against Drools 5.3.0). The Broker example code in both
versions is identical, so only the Drools-internal code changes matter.
When running the 5.2.0 Broker demo to the end, you get one
java.text.ParseException (given the structure of the example code, that's
expected, albeit not elegant, and not the focus of our investigation). In
particular, no matter how many lines stocktickstream.dat contains, you
always get exactly one ParseException at the end.
Contrast this with running the 5.3.0 Broker demo: At the end you get N
occurrences of java.text.ParseException, where N is the number of lines in
stocktickstream.dat. So for 14 lines you get 14 occurrences of
ParseException. Looking at two specific methods shows us why:
Method
org.drools.examples.broker.events.EventFeeder.FeedJob.execute(JobContext):
public void execute(JobContext context) {
this.sink.receive( ((FeedContext) context).event );
if ( this.source.hasNext() ) {
((FeedContext) context).setEvent( this.source.getNext() );
this.trigger.setNextFireTime( ((FeedContext)
context).getEvent().getDate() );
clock.scheduleJob( this,
context,
trigger );
}
}
Note in particular how this method already takes care of scheduling the
next job execution by updating the next fire time of the job's existing
FeedTrigger instance. Unfortunately, in Drools 5.3.0,
DefaultTimerJobInstance.call() does a duplicate scheduling of the same job:
Method org.drools.time.impl.DefaultTimerJobInstance.call():
public Void call() throws Exception {
this.trigger.nextFireTime(); // need to pop
if ( handle.isCancel() ) {
return null;
}
this.job.execute( this.ctx );
if ( handle.isCancel() ) {
return null;
}
// our triggers allow for flexible rescheduling
Date date = this.trigger.hasNextFireTime();
if ( date != null ) {
scheduler.internalSchedule( this );
}
return null;
}
So, every job is duplicated and that's why there are 2*N calls to
org.drools.examples.broker.events.StockTickPersister.load() instead of N.
I think the rescheduling inside DefaultTimerJobInstance.call() is
incorrect. For one, it breaks backward compatibility, and it is unexpected.
The job should be in charge of deciding whether there is another job to
schedule or what to do. Implicitly scheduling the next job just by updating
the trigger time is a little too much magic.
Issue F2:
This is the bug that causes my unit tests to fail. I have not pinpointed
the root cause, but it seems to have to do with the event scheduling Drools
does as part of its job execution mechanism. Its symptom is a
NullPointerException during insertion of an event. What makes it so tricky
is that with the out-of-the-box configuration, Drools catches such
exceptions in org.drools.time.impl.PseudoClockScheduler.runCallBacks() and
passes them to the aptly named "DoNothingSystemEventListener", which
literally does nothing, not so much as logging (the methods are empty). So
you don't actually know that the event insertion failed, you only wonder
why your mock WorkingMemoryEventListener is telling you that your
expectations are not met. The stack trace (as copied from the Eclipse Debug
view, hence the unusual formatting) inside Drools is:
Date.getMillisOf(Date) line: 939
Date.compareTo(Date) line: 959
DefaultTimerJobInstance.compareTo(DefaultTimerJobInstance) line: 38
DefaultTimerJobInstance.compareTo(Object) line: 13
PriorityQueue<E>.siftUpComparable(int, E) line: 582
PriorityQueue<E>.siftUp(int, E) line: 574
PriorityQueue<E>.offer(E) line: 274
PriorityQueue<E>.add(E) line: 251
PseudoClockScheduler.internalSchedule(TimerJobInstance) line: 136
PseudoClockScheduler.scheduleJob(Job, JobContext, Trigger) line: 126
ObjectTypeNode.assertObject(InternalFactHandle, PropagationContext,
InternalWorkingMemory) line: 230
EntryPointNode.assertObject(InternalFactHandle, PropagationContext,
ObjectTypeConf, InternalWorkingMemory) line: 244
NamedEntryPoint.insert(InternalFactHandle, Object, Rule, Activation,
ObjectTypeConf) line: 330
NamedEntryPoint.insert(Object, boolean, boolean, Rule, Activation) line: 291
NamedEntryPoint.insert(Object) line: 116
NamedEntryPoint.insert(Object) line: 48
<my code calling into Drools>
Here is method
org.drools.time.impl.DefaultTimerJobInstance.compareTo(DefaultTimerJobInstance):
public int compareTo(DefaultTimerJobInstance o) {
return this.trigger.hasNextFireTime().compareTo(
o.getTrigger().hasNextFireTime() );
}
Essentially, this method calls java.util.Date.compareTo(Date) with a null
argument, which, as documented, causes a NullPointerException. Sometimes,
this.trigger.hasNextFireTime() already returns null, and then the NPE gets
thrown in DefaultTimerJobInstance.compareTo() itself.
I've seen different stack traces leading to this NPE, so this must be
affecting scheduling and job execution quite broadly.
Issue F3:
It is debatable whether this is a bug, but it is a backward-compatibility
breaking change. Previously, when scheduling a job with
org.drools.time.TimerService.scheduleJob(Job job, JobContext ctx, Trigger
trigger) (for both real-time and pseudo clock), you could pass a null
JobContext (say, because you didn't need one), and it would work. However,
in Drools 5.3.0, this causes a NullPointerException at:
org.drools.time.impl.DefaultTimerJobFactoryManager.createTimerJobInstance(Job,
JobContext, Trigger, JobHandle, InternalSchedulerService) line: 25
I realize that if it's not in knowledge-api-<version>.jar, it's not an
official API, but the available interfaces and classes in
org.drools.time.** (as used in the Broker example) are *very* useful for
test harnesses *and* for production code (for implementing dynamic timers,
for instance). So, this is more of a heads-up: If you are suddenly getting
an NPE, this might be the cause.
Please let me know whether I should create JIRA bug reports for issues F1
and F2. Also, I'd be interested to hear whether others have run into issues
with Fusion in Drools 5.3.0.