RFC: Overriding @Provides methods considered harmful

[Luke Sandberg]

While preparing a recent change to how @Provides methods are invoked, I came across some confusing behavior when @Provides methods are overridden.  The initial issue i saw was for a case like this:

class Super extends AbstractModule {

  @Override void configure() {}

  @Provides Number provideNumber() {

    return 1;

  }

}

class Sub extends Super {

  @Provides Integer provideNumber() {

    return 2;

  }

}

Injector injector = Guice.createInjector(new Sub());

assertEquals(2, injector.getInstance(Number.class));

assertEquals(2, injector.getInstance(Integer.class));

This happens because Sub.provideNumber is a covariant override of Super.provideNumber.  So 2 provider methods are bound, but both call Sub.provideNumber. 

If instead you wrote:

class Sub extends Super {

  @Provides Number provideNumber() {

    return 2;

  }

}

then Guice.createInjector will throw a CreationException for duplicate bindings for Key.get(Number.class).

There are a lot of different combinations for how you could override an @Provides method depending on whether or not you mark the override with @Provides and whether or not you change the key for the method.  This gets especially weird if you start adding scoping annotations into the mix (e.g. if the override is marked @Singleton but the parent isn't).  Currently, the Guice implementation of @Provides doesn't do anything special to try to detect these cases.

Proposal:

My proposal is to make it an Error to override an @Provides method in all cases. Guice will essentially enforce that @Provides methods are private/final (whether or not they are actually declared as such).

The goal of this proposal is to eliminate opportunities for confusion and to make modules easier to understand.

FAQs

Q: What about users who overrides @Provides methods for testing purposes?

Guice has other mechanisms to override bindings (e.g. Modules.override), or you can structure your test so that the parent @Provides method isn't installed, maybe by splitting it into a separate Module class that you don't install in the test.  Or if you really want to use method overriding you can use a template method pattern and just override a different method and have the @Provides method call that one.

Q: What about migration of legacy code that uses this feature?

I'm not sure yet.  Guice has a some precedent for using java system properties for feature flags, so we may introduce a system property ('guice_allow_provides_overrides'?) to enable/disable this error.

[Christian Gruber]

+1. I would even consider making it a warning if it isn't marked final,
to encourage people to add those constraints.

There's an unfinished (it works, but is not sufficiently tested)
annotation processor for some guice static analysis, this could also be
implemented there (and have it enabled) so that some of these runtime
errors can be caught at compile-time. (It should remain a runtime error
as well in case people do not run the annotation processor / validator).
If someone does need the old behavior, they may mark the error
suppressed.

The legacy flag is not pretty, but it's certainly a viable option,
though here also I would spout a warning on startup.

Christian.

On 23 May 2014, at 12:20, Luke Sandberg wrote:

> While preparing a recent

> change<https://code.google.com/p/google-guice/source/detail?r=409e0f578b620c38f6c8626dee78783219d2956e>

> --
> You received this message because you are subscribed to the Google
> Groups "google-guice" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to google-guice...@googlegroups.com.
> To post to this group, send email to google...@googlegroups.com.
> Visit this group at http://groups.google.com/group/google-guice.
> For more options, visit https://groups.google.com/d/optout.


Christian Gruber :: Google, Inc. :: Java Core Libraries :: Dependency
Injection
email: cgr...@google.com :::: mobile: +1 (646) 807-9839

[Luke Sandberg]

On Friday, May 23, 2014 4:20:30 PM UTC-7, Christian Gruber wrote:

+1.  I would even consider making it a warning if it isn't marked final,
to encourage people to add those constraints. 

 I'm slightly negative on forcing 'final' modifiers (purely due to boilerplate issues), though I'm sure i could be convinced.

There's an unfinished (it works, but is not sufficiently tested)
annotation processor for some guice static analysis, this could also be
implemented there (and have it enabled) so that some of these runtime
errors can be caught at compile-time. (It should remain a runtime error
as well in case people do not run the annotation processor / validator).
  If someone does need the old behavior, they may mark the error
suppressed.

 

The annotation processor sounds awesome.  If it is just a matter of writing tests I would be willing to offer some help to get that out the door :)

The legacy flag is not pretty, but it's certainly a viable option,
though here also I would spout a warning on startup.

Since all these issues would be detected at configure time, we could probably just log warnings/severes if the error is disabled.  I agree that adding a flag for this is kind of gross, but i assume it is neccesary for backwards compatibility.  Though i don't know what the policy is for breaking changes like this.  I don't think that this is a large issue (based on a cursory review of googles code base), but still it could definitely break users.

[Christian Gruber]

Hey Steven - you had the processor in a git repo, right? I haven't had
a chance to fully follow up on it, but we have takers to help get it all
tested. Let's coordinate and get the code into some place where it can
be worked on and tested.

As to breaking changes, we already have some breaking changes if I
recall in Guice 4, so it might be timely (though folks on this list
should pipe up as to whether that's going to rain doom upon them. :)

Christian

>>> an email to google-guice...@googlegroups.com <javascript:>.

>>> To post to this group, send email to

>>> google...@googlegroups.com<javascript:>.

>>
>>> Visit this group at http://groups.google.com/group/google-guice.
>>> For more options, visit https://groups.google.com/d/optout.
>>
>>
>> Christian Gruber :: Google, Inc. :: Java Core Libraries :: Dependency
>> Injection

>> email: cgr...@google.com <javascript:> :::: mobile: +1 (646) 807-9839

[Luke Sandberg]

FYI: I have submitted this change to google's internal copy of guice and it should be present at HEAD whenever the repo sync next happens.

any update on the annotation processor?

To unsubscribe from this group and stop receiving emails from it, send an email to google-guice+unsubscribe@googlegroups.com.

To post to this group, send email to google...@googlegroups.com.
Visit this group at http://groups.google.com/group/google-guice.
For more options, visit https://groups.google.com/d/optout.

[Sam Berlin]

(The changes are now pushed, along with a few other changes.)

To unsubscribe from this group and stop receiving emails from it, send an email to google-guice...@googlegroups.com.

To post to this group, send email to google...@googlegroups.com.
Visit this group at http://groups.google.com/group/google-guice.

To view this discussion on the web visit https://groups.google.com/d/msgid/google-guice/CAO9V1MKOtQn7UhMfMHTY8PwWbO6baBVhBG-TDk-3zMt44UOdFA%40mail.gmail.com.