ProxyMetaClass doesn't work as expected with StubFor/Mockfor

Added confluence code tags

I am not sure the above is true for Groovy objects which each get their own metaClass. For Java objects which share a metaClass it would be true I think.

The above script was executed with groovyConsole. I've attached a simple testcase and ran it againts the current HEAD (9124) and it failed, as you may notice the CUT is a GroovyObject.

I can replicate the issue that you have raised but also observe that behaviour in 1.0 and thinking a bit more about it am not sure the current behaviour isn't the correct one. Happy to be proved wrong though - we certainly should be certain of what the correct bahaviour should be.

Well if any instance of the class being mocked shares the same MetaClass while inside use() then we have a bug (this is just by looking at the base level)
Now at the upper level, just looking at how mocks should behave, does it make sense that only instances of said class should share the same MetaClass if they are created inside use() (the current behavior)?

To give you more insight, I wanted to see if OGB + Mocks are a good idea to create test data from a non-anemic domain model, but because OGB relies on property setters and those methods were not mocked I got into the following scenarios:

a) test model graph is created outside use() so any mocked instance does not have any expectation.
b) test model graph is created inside use(), as setters are not demanded the test will fail while creating the graph.

OK, I can see your issue now. It may pay to try to come up with a minimalistic testcase involving OGB. I was hoping to add more test cases myself in this area over the next few days around using mock instances but wasn't planning to involve OGB. Would you need support for property setters/getters even after this fix is in place?

Attached new version of testcase with two OGB tests.

The story is that Groovy objects initialize metaClass field in
constructor. So if you create your Groovy object outside stub closure
it has regular metaclass instead of mocked one.

I really don't know what the right behaviour should be.

In theory, we can change compiler and runtime to do some check in
getMetaClass () that metaclass stored in Groovy object was replaced in
the registry by another one and reassign stored value.

What I really don't like is the design where one thread replace
metaclass to do testing with mock and that effect all threads
executing objects of the class.

I think that for the case of stub/mocks the proxyMetaClass should be changed for all instances of the mocked class regardless when they were instantiated while they are under the influence of the use closure, in the same way that categories do. This means we forfeit multi thread testing with stubs/mocks and the metaClass will only be available in the current change for the duration of the use closure, once the closure finishes, the previous metaClass is restored in the registry. This will cover the A scenario I outlined.

Now on the B scenario, what's happening is something i did no expect coming from using JMock's stubs, which is that the Company.getEmployees() was not demanded thus throwing an AFE. It seems like StubFor is a weaker version of MockFor (does not care of record sequence) but stronger than JMock's stubs (which will let you call any method without cardinality check). What I would love is have the option for stubs to be behave like partial suppliers of behavior when used with concrete classes and providers of default values when used with interfaces/abstract classes; something different of what StubFor provides now.

Meaning that the following scenarios may be possible:

interface MyInterface {
   boolean doSomething()
   Object doAnotherThing( String param )
}

class CUT {
   MyInterface collaborator
   def doit() { collaborator.doSomething() }
}

def stub = new GroovierStub(MyInterface)
// configure demands if any
stub.use {
   // instantiate the CUT which will use instances of MyInterface
   def cut = new CUT()
   assert !cut.doIt()
   // because the default impl for doSomething() would be to return false
   // as it is the default value for a boolean
}

When stubbing a concrete you may override any method you'd like, those that were not overridden will still have their normal behavior. Stubbing an abstract class follows a mix of concrete/interface where concrete methods may or may not be overridden with demands (as in concrete, retaining their behaviors those who were not overridden) and default return values for those abstract methods that were not demanded.

In short, this is something that StubFor can't provide unless it breaks compatibility, and if something like this proceeds it should be done after 1.1 ships.

</rant>

I have submited test, which shows how to achieve behavior you need. The trick is that for groovy objects you have to use stab.use(object)

Can I check everyone's understanding for what the situation is for 1.1?

Using this base class:

class Dice {
    int roll() { 6 }
}

If you instantiate a Groovy object outside the use and don't use the object version of use, the outside instantiations won't be stubbed/mocked, e.g.:

def diceStub = new StubFor(Dice)
private count = 1
diceStub.demand.roll(0..10) { count++ }
def d1 = new Dice()
println d1.roll() // => 6
diceStub.use {
    def d2 = new Dice()
    println d1.roll() // => 6
    println d1.roll() // => 6
    println d2.roll() // => 1
    println d2.roll() // => 2
    println d2.roll() // => 3
}

If you use the object version of use, that object will be stubbed/mocked but others won't, e.g.:

diceStub = new StubFor(Dice)
count = 1
diceStub.demand.roll(0..10) { count++ }
def d3 = new Dice()
println d3.roll() // => 6
diceStub.use(d3) {
    def d4 = new Dice()
    println d3.roll() // => 1
    println d3.roll() // => 2
    println d4.roll() // => 6
    println d4.roll() // => 6
    println d4.roll() // => 6
}

If you want to mock/stub multiple objects of the same type you would need nested use statements with the demands representing the demands across all objects, e.g.:

diceStub = new StubFor(Dice)
count = 1
diceStub.demand.roll(4..4) { count++ }
def d5 = new Dice()
println d5.roll() // => 6
diceStub.use(d5) {
    def d6 = new Dice()
    println d6.roll() // => 6
    diceStub.use(d6) {
        println d5.roll() // => 1
        println d5.roll() // => 2
        println d6.roll() // => 3
        println d6.roll() // => 4
    }
}

Have we got a potential fix for making this work as expected?

I'm moving it back to 1.2, not 2.0.
And if there's a potential fix for 1.1, we may potentially put it it back in 1.1.

Thanks Alex, but as Paul noted 'stub.use(object)' will only work for object and not other instances of the same class. I still think the category approach (anything that is under the influence of the use closure and matches the stubbed class) is the way to go though in terms of how it should be implemented poses the question on how to handle multi threading.

Because of this I wouldn't want to rush things to get it into 1.1 as now what we have discussed so far makes me think this may not be exactly a bug but a false expectation on how StubFor should work (though if the ProxyMetaClass is added to the registry then any instance regardless of its creation time should be affected by it during the use closure, right? at least that is what I gather).

The nested use(object) work but seems a little un-Groovy by the way.

Not helpful for the final solution of this issue but as a note, you can now use "instance" meta mocks as well as "shared" meta mocks as follows:

diceStub = new StubFor(Dice)
diceStub.demand.roll() { 3 }
diceStub.demand.roll() { 4 }
def d1 = diceStub.proxyInstance()
def d2 = diceStub.proxyInstance()
def d3 = diceStub.proxyInstance()
println d1.roll() // => 3
println d1.roll() // => 4
println d2.roll() // => 3
println d3.roll() // => 3
println d3.roll() // => 4
println d2.roll() // => 4

Wondering whether this should be closed. Current behavior:

import groovy.mock.interceptor.StubFor

class Speaker {
   String name
   def startLecture() { "Starting..." }
}

def stub1 = new StubFor(Speaker)
stub1.demand.startLecture() { "Intercepted!" }
def stub2 = new StubFor(Speaker)
stub2.demand.startLecture() { "Intercepted!" }
def speaker1 = new Speaker()
stub1.use {
   def speaker2 = new Speaker()
   assert speaker1.startLecture() == "Starting..."
   assert speaker2.startLecture() == "Intercepted!"
}
stub2.use(speaker1) {
   def speaker2 = new Speaker()
   assert speaker1.startLecture() == "Intercepted!"
   assert speaker2.startLecture() == "Starting..."
}

The behavior while not perfect is now reasonably documented and well-known. In some ways it seems a reasonable approach. The metaClass is something which seldom changes hence it might be argued that once an instance (created outside the 'use') has one it should be left untouched. The more frequent use of EMC versus traditional meta-programming makes this less clear though.

Or if not closed, should we set fix version to 2.0?

I would say "close" (current status okay for me), but I'm okay either way.
We could indeed revisit things in 2.0, if a new MOP can help further, so that's why moving to 2.0 is not a bad idea either

Marking it as resolved per previous comments.