Hi Marten, Sorry for the late response, it's been a busy week. Thanks for working on this, I'll try to have a look at your branch this weekend.
In the meantime I'll leave some comments below. > On Feb 12, 2015, at 22:43, Marten Kenbeek <[email protected]> wrote: > > Current status: > • Models in general seem to be working. > • Trying to override a concrete (or explicitly locked) field with an > abstract field will keep the concrete field, but not raise any errors. This > occurs when inheriting from (AbstractBase, ConcreteBase). > • I might still change this to raise an error, it migth cause a > few headaches when fields don't appear while they're first in the mro. This should definitely be an error. We recently added checks so two concrete parents can't have clashing fields, this is similar. > • Trying to override field.attname (e.g. foo_id) will fail checks > unless you explicitly set foo = None. > • Trying to override an abstract field with a reverse relation will > fail checks unless you explicitly set foo = None. That's good. > What is still needed: > • ? More extensive model tests. During development a few more bugs > popped up. I've written tests for them, but there might be some more. > • Explicit migration tests. Existing migration tests pass, but don't > include the new scenario's. > • Documentation. > My WIP branch can be found at > https://github.com/knbk/django/compare/ticket_24305. > > Op dinsdag 10 februari 2015 14:36:53 UTC+1 schreef Marten Kenbeek: > Hi Aron, > > With option 3, this would work: > > class Cousin(models.Model): > child = models.ForeignKey(Child, allow_override=True) > > but it would raise an error without `allow_override`. > > I think my question really is: should we treat reverse relations as > first-class fields, or not? If we do, 1) would be the logical choice but can > cause problems. 3) would deal with those problems in a way and still allow a > relation field from another model to override a local field. If we don't, > that kind of implies 2). Removing the field with `None` would implicitly > allow a reverse relation to take its place, but that reverse relation in and > of itself would not be allowed to override any abstract fields. > > This is just an example. In real code this would more likely happen with an > explicit `related_name`. Renaming the related name and living with an unused > `cousin_set` attribute is the current solution, but with this feature > proposal it would be nice to also have a way to override the inherited field > and use it for a reverse relation instead. > > > Op dinsdag 10 februari 2015 02:28:58 UTC+1 schreef Aron Podrigal: > Why should this be treated differently than the general behavior when > realted_names clash when you have more than one foreign key to the same > model? So as one would normally do > set the related_name explicitly to some other value. > setting the field to None is just the way of removing a field and has nothing > more special related to the auto created reverse field descriptor. > about option 3 I didn't quite understand. can you explain a bit more? +1 to what Aron said. > > On Monday, February 9, 2015 at 4:25:22 PM UTC-5, Marten Kenbeek wrote: > I'd like some additional opinions on the following: > > Say you have the following classes: > > class Parent(models.Model): > cousin_set = models.CharField(max_length=100) > > class Meta: > abstract = True > > class Child(Parent): > pass > > class Cousin(models.Model): > child = models.ForeignKey(Child) > > Obviously, the `cousin_set` field inherited from `Parent` clashes with the > reverse relation from `Cousin`. > > I can see the following options: > 1) Allow the reverse object descriptor to overwrite the field. No this would be very confusing. > 2) Only allow the reverse object descriptor to overwrite the field if it is > explicitly set to None on the target model. Yes but as Aron pointed out we aren't specifically "allowing" the descriptor to override the field, it's just that the field was already removed (through None) so the reverse descriptor isn't clashing with anything. > 3) Only allow the reverse object descriptor to overwrite the field if the > foreignkey/m2m/o2o field itself has a flag set to explicitly override. Let's not do that, it's clunky, I also don't like the locking mechanism mentioned above. > 1) is consistent with the behaviour of local fields, but I think it will be > problematic if other models can silently overwrite a field. 3) would still > allow other models to affect local fields, but at least it has a fail-safe > that prevents you from accidentally overriding fields. 2) would only allow > the inheriting model itself to change which fields it inherits. > > Problems caused by option 1) would be hard to debug when you don't know which > code overrides your field, so I wouldn't do that. I think 2) would be the > cleanest and most consistent way. Only local fields would override parent > fields, but the sentinel value `None` would remove the field and free the > name for reverse relations. I can also see the advantage of 3) over 2) when > you don't have access to the model on the other side. However, I don't know > enough about foreign key internals to know if 3) is at all feasible. What > happens e.g. when only the target is loaded in a migration? Would it pick up > that the remote foreign key overrides a local field? As adding reverse > relations is a lazy, or at least delayed operation afaik, would it still be > save to rely on that to override fields? > > I believe my current plans for the patch would automatically create situation > 2 without any extra work. The field would no longer exist on the child class > when the reverse relation is added. Option 3) would require an additional > patch to the code that adds the reverse relationship, but it allows for some > extra options. > > Any input? Additional options are also welcome. > > Op zondag 8 februari 2015 21:09:41 UTC+1 schreef Marten Kenbeek: > The general reaction seems to be a yes if we can work out the kinks, so I > went ahead and created a ticket: https://code.djangoproject.com/ticket/24305 > > @Aron That's a good question. One option would be to lock certain fields, so > that they can't be changed if they are an integral part of the model. That > would be a simple solution, but that won't help for existing code that > doesn't lock the fields. It won't break existing code, but it won't protect > for errors either. The opt-in version (i.e. an 'unlock' attribute) would lock > many fields which would otherwise be completely safe to overwrite. > > Another option would be more elaborate "requirements" for a manager or some > methods, i.e. allow the manager to specify the necessary class of a certain > field or a minimum length. If the modeldoesn't meet the requirements, the > manager or some of the methods will not be inherited. While it allows for > more control, this option would greatly increase the complexity of the patch > and requires more from the developers of custom managers. It can also cause > issues when the requirements aren't up-to-date with the manager's methods. > > We could also say that it is the users responsibility and don't provide > special protection, in line with the fields themselves, but I guess that this > would generally be more problematic for custom managers. It can also cause > silent bugs when the manager's methods don't work as intended but won't raise > an exception either, which is not a good idea imho. > > I think the locking approach would be the easiest and most pragmatic method. > I think it's still - in part - the users responsibility to confirm that a > field can be overridden. The Django documentation could, where applicable, > document the requirements on fields that can be overridden, i.e. that an > AbstractUser's username must be a CharField (which isn't necessarily true, > just an example). > > @Loïc The bases are indeed traversed in the reversed order. > > Op zondag 8 februari 2015 08:19:57 UTC+1 schreef Loïc Bistuer: > That's what we've done in Django 1.7 for Form/ModelForm (#19617, #8620), and > I completely agree that it should be possible to shadow fields from abstract > models. The docs even suggest that this may be possible in the future; > quoting the relevant section: "this [shadowing] is not permitted for > attributes that are Field instances (at least, not at the moment)." > > For consistency with forms - and because we've put a great deal of thinking > into it - the implementation should be: you can shadow a field with another > field, or you can remove a field using None as a sentinel value (see #22510 > for more details). > > I believe we have a safety net in the form of migrations, if you accidentally > shadow a field, migrations will pick it up, so it's unlikely to go unnoticed. > > I'd be quite happy to shepherd this patch. > > Unit tests should cover those scenarios: > > - An abstract model redefines or removes some fields from a parent abstract > model. > - A concrete model redefines or removes some fields from a parent abstract > model. > - Ensure that the standard MRO resolution applies when you do > Concrete(AbstractA, AbstractB) with AbstractA redefining a field from > AbstractB. > > The last case may be tricky because if I remember well ModelBase iterates > through the MRO in the wrong direction (i.e. bases should be iterated over in > reverse). > > We also need some tests to show how migrations behave. > > -- > Loïc > > > On Feb 7, 2015, at 09:33, Marten Kenbeek <[email protected]> wrote: > > > > Hi Russ, > > > > I can see your point on accidentally overriding fields, though I'm not sure > > I agree. In any other case, such as normal attributes and methods, there is > > no safety net for overriding attributes either. Any model that would be > > affected by this change would also raise an error on import without the > > patch, so existing functional code won't be affected. On the other hand, a > > new parameter for the field wouldn't be that much of a hassle to implement > > or work with. I'd be interested to hear what others think about this. > > > > There are more than a few questions on stack overflow that expect this > > behaviour, even if the docs specifically mention that it won't work. If > > users intuitively try to override fields in this manner, I think it would > > be reasonable to allow this without any explicit arguments. > > > > We can always restrict what you can override, at least as the default > > behaviour, e.g. so that you can only use subclasses of the original field. > > That would make code that depends on the abstract field less prone to bugs, > > though subtle bugs can still happen if you e.g. override the max length of > > a field. > > > > This was indeed just a proof-of-concept. That remark was meant more like > > "it doesn't appear to break everything". > > > > Marten > > > > Op vrijdag 6 februari 2015 23:48:55 UTC+1 schreef Marten Kenbeek: > > Hey, > > > > While fiddling with django I noticed it would be trivial to add support for > > shadowing fields from abstract base models. As abstract fields don't exist > > in the database, you simply have to remove the check that reports name > > clashes for abstract models, and no longer override the field. > > > > This would allow you to both override fields from third-party abstract > > models (e.g. change the username length of AbstractUser), and to override > > fields of common abstract models on a per-model basis. Previously you'd > > have to copy the original abstract model just to change a single field, or > > alter the field after the model definition. The first approach violates the > > DRY principle, the second approach can introduce subtle bugs if the field > > is somehow used before you changed it (rare, but it can happen). > > > > This patch adds the feature. It seems to work correctly when using the > > models and creating/applying migrations, and passes the complete test > > suite. > > > > What do you guys think about this feature? > > > > -- > > You received this message because you are subscribed to the Google Groups > > "Django developers (Contributions to Django itself)" group. > > To unsubscribe from this group and stop receiving emails from it, send an > > email to [email protected]. > > To post to this group, send email to [email protected]. > > Visit this group at http://groups.google.com/group/django-developers. > > To view this discussion on the web visit > > https://groups.google.com/d/msgid/django-developers/072149e1-c794-446d-957b-f5fc5df87096%40googlegroups.com. > > > > For more options, visit https://groups.google.com/d/optout. > > > -- > You received this message because you are subscribed to the Google Groups > "Django developers (Contributions to Django itself)" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at http://groups.google.com/group/django-developers. > To view this discussion on the web visit > https://groups.google.com/d/msgid/django-developers/6e3e363f-bdc6-410b-b145-76085051e053%40googlegroups.com. > For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/django-developers. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/18C7DC67-DBD7-417E-905B-AF1A13777C0D%40gmail.com. For more options, visit https://groups.google.com/d/optout.
