document that ManyToMany fields can't be in unique_together

[hugo]

From IRC:

• hugo- any way to get a unique_together that works with ManyToMany fields?
• adrian_h I don't believe so, at the moment
• hugo- yeah. I had it working with some simple validator, until I realized that my validator can't know wether I am creating or updating (and if updating, which one I am updating) :-)
• hugo- I _think_ it could be done with the same check you do in manipulators_validator_unique_together in core.meta for ManyToOne
• hugo- or at least a very similar one
• adrian_h If a model has unique_together = (('slug', 'sites'),), where sites is a manytomany field, would the validator check that the slug is unique per each individual site, or per the specific combination of the sites?
• hugo- per each site
• hugo- at least that's how you use it in reverse: you use sites.get_current() to get the curent site and then for example check get_current().get_flatfile_object() to find the flatfile for that url for that site
• hugo- currently you can add flatfiles for the same site with the same URL and get a nice exception ...
• adrian_h Good point
• adrian_h This is worth a ticket

[mrts]

This should be provided by model validation. As model validation is scheduled as "maybe" for 1.0, marking as 1.0 beta.

[Jacob]

Model validation is getting pushed until post-1.0, so this is, too.

[(none)]

Milestone post-1.0 deleted

[lpetrov@…]

any future plans for this ticket ?

[Preston Holmes]

This should be in the db/orm component

for reference an attempt at implementation would have to be made here:

django.db.models.base.Model._perform_unique_checks

If the implementation is not practical (this ticket closed as wontfix) a note of the limitation should be noted in the docs

[Carl Meyer]

I don't think we should allow m2m fields to be listed in unique_together.

In practical terms, this is cross-model validation between two models (the main model and the m2m through model, which may be implicit). This is quite challenging and different from a normal unique_together, because the validation code for the main model runs before the m2m models have even been saved, so implementing this requires either listening to an m2m_changed signal, or putting the validation code onto the m2m through model. In a regular unique_together, there is an actual UNIQUE constraint created in the database covering the unique_together fields; with an m2m unique-together it could only be enforced by the model-validation code. IOW, the underlying implementation would be quite different from the implementation of a normal non-m2m unique_together, and those differences would be bound to leak up into application code.

Also, the meaning of listing an m2m field in a unique_together is not self-evident, as demonstrated by the IRC conversation in the description on this ticket. If "charfield" is unique_together with "m2mfield", does that mean charfield must be unique with respect to each individual value of m2mfield, or with respect to the combination of m2mfield values? Clearly in the FlatPage case the former is desired, but I can certainly imagine cases in which the latter would be desired.

Lastly, there's already a sort of precedent here - ManyToManyFields cannot be set as "unique" (presumably because it would result in similar semantic questions and implementation challenges). It seems pretty reasonable that a field type which can't be set as unique also can't be included in unique_together.

In light of all this complexity, I think validation of uniqueness related to m2m fields should be left up to the application developer, via signals or an explicit through model, rather than built in to the ORM. I'm changing the component here to documentation, and the summary to reflect that we need to document that m2m fields can't be included in unique_together.

[Dan Poirier]

Doc update in pull request ​https://github.com/django/django/pull/77
​https://github.com/django/django/pull/77.diff

[Tim Graham]

In [17314]:

Fixed #702 - Documented that ManyToMany fields can't be in unique_together; thanks poirier for the patch.

[Tim Graham]

In [17315]:

[1.3.X] Fixed #702 - Documented that ManyToMany fields can't be in unique_together; thanks poirier for the patch.

Backport of r17314 from trunk.