Add support for multi-column primary keys.

[Jacob]

In the interest of being able to use Django as an admin interface for any db schema, it really should support multiple column primary keys.

Currently, you can "fake" it by declaring one of the keys to be primary in Django and adding a unique constraint to the model. You'd have to do without the auto-generated SQL that Django gives you, but if you're smart enough to know how and why to use multiple primary keys, you can probably write the schema by hand. Still, this is less than ideal.

This is likely to be very difficult.

[Jacob]

This is something that gets asked about a lot, so I'm going to reopen this ticket as a place to keep track of it.

Personally I don't have time to work on this, but if the LazyWeb can work up a patch, it would be quite nice to have. From an email I wrote a while back, here are the issues (that I'm aware of) that would need to be solved to make this work:

There's three basic problems in dealing with composite primary keys in Django.

The first is that a number of APIs use "obj._meta.pk" to access the primary key field (for example, to do "pk=whatever" lookups). A composite PK implementation would need to emulate this in some way to avoid breaking everything.

Second, a number of things use (content_type_id, object_pk) tuples to refer to some object -- look at the comment framework, or the admin log API. Again, a composite PK system would need to somehow not break this.

Finally, there's the issue of admin URLs; they're of the form "/app_label/module_name/pk/"; there would need to be a way to map URLs to objects in the absence of a primary key.

[brantley]

It seems to me that multiple primary keys could be represented simply as tuples. The admin could function as: "/app_label/module_name/key1,key2/"

[Adrian Holovaty]

Brantley: In your URL example, what happens when the key value contains a comma?

[anonymous]

Adrian: what if in the current setup, a primary key has a "/" in it?

[Adrian Holovaty]

The current setup works with primary keys with slashes in 'em, because the admin-site regular expression is greedy. Try it yourself -- and if there's a problem, please do let us know.

[brantley]

I think that might be a fringe case that could be overcome in many ways.

[anonymous]

It also seems like the admin can't recognize any unique_together foreign key pair.

[C8E]

What about /app_label/module_name/key1/key2/? That way we can think to access to partial key query /app_label/module_name/key1/, like in the date based generic views.

[anonymous]

in regards to: "/app_label/module_name/key1,key2/". though not as efficient, primary keys can be text. what happens if they contain a comma, where pk1 = "i do, sometimes" and pk2 = "value of key 2"? /app_label/module_name/i do, sometimes,value of key 2/. or would this not happen?

[Gary Wilson <gary.wilson@…>]

Seems like the details for this still need to be decided.

[buriy <yuri@…>]

so why not do escaping for such keys?
Add class ComplexKey...
add some str method or something like this for getting escaped representation of the key.
and what's the best escape sequence for ','... please discuss that without me:)

[anonymous]

Replying to buriy <yuri@buriy.com>:

and what's the best escape sequence for ','

A second comma?

from django.db.vapourware import urlencode_pk
assert urlencode_pk(("ab","c") == "ab,f"
assert urlencode_pk(("a,b","c") == "a,,b,f"
assert urlencode_pk(("a,,b","c") == "a,,,,b,f"
assert urlencode_pk(("i do, sometimes","value of key 2") == "i do,, sometimes,value of key 2"

...although I'm trying to think of a simple RE that would code this, and I can't think of anything that would work.

[ilatypov]

Does the issue of encoding a composite key into URL prevent from solving this ticket?

To address the URL encoding issue, may I suggest to consider comma an unsafe character when encoding parts of the composite primary key? Then an unprotected comma can serve a clear mark of part separator. According to the URI syntax at W3C web site,

​http://www.w3.org/Addressing/URL/5_URI_BNF.html ,

the comma character is allowed to appear unprotected in URIs.

So encoding a composite primary key should be as simple as

pk = [url_encode(pk_part, more_unsafe=",") for pk_part in pk_list]
url = ",".join(pklist)

The major roadblock is still the issue of supporting composite primary keys in the models.

[ilatypov]

I meant ",".join(pk).

[David Cramer]

I'm currently working up some support for this -- the work will depend and need to wait on the queryset refactoring code.

The currently plan is to maintain backwards compatibility on instance._meta.pk, but it will throw an exception if accessed when there are multiple pks, and it should be deprecated. There will be a new attributes, pks on the meta class, which will return a tuple of the primary key fields.

[ulvinge@…]

Sorry if this isn't the right place to post this, but I'm a noob... I've only used Python and Django for a week...

Currently, you can "fake" it by declaring one of the keys to be primary in Django and adding a unique constraint to the model. You'd have to do without the auto-generated SQL that Django gives you, but if you're smart enough to know how and why to use multiple primary keys, you can probably write the schema by hand. Still, this is less than ideal.

No you really can't "fake" it... You can't even with custom sql do it.

The problem is in the save function, because when you save an object it overrides all values with the declared primary key.

An example:

class book():
    shelf  = IntegerField(primary_key=True)
    level  = IntegerField()
    index  = IntegerField()

    class Meta:
        unique_together = ('shelf', 'level', 'index')
    # don't mind the following line right now
    primary = ('shelf', 'level', 'index')

Currently you can only have one book per shelf with the current django code... (I actually use 0.96, but I don't think it has changed...)

That was the problem, now to the solution! (Note, as I said, I'm a noob, and I just hacked an afternoon and came up with this)

It needs to have a field called primary specified as shown in the above model.

import django.db.models.manipulators 
import django.db.models.manager 
from django.core import validators 
from django.core.exceptions import ObjectDoesNotExist 
from django.db.models.fields import AutoField, ImageField, FieldDoesNotExist 
from django.db.models.fields.related import OneToOneRel, ManyToOneRel 
from django.db.models.query import delete_objects 
from django.db.models.options import Options, AdminOptions 
from django.db import connection, backend, transaction, models 
from django.db.models import signals 
from django.db.models.loading import register_models, get_model 
from django.dispatch import dispatcher 
from django.utils.datastructures import SortedDict 
from django.utils.functional import curry 
from django.conf import settings 
from itertools import izip 
import types 
import sys 
import os 
 
class Model(models.Model): 
    def save(self): 
        dispatcher.send(signal=signals.pre_save, sender=self.__class__, instance=self) 
 
        non_pks = [f for f in self._meta.fields if not (f.primary_key or (f.name in self.primary))] 
        otpk = [f for f in self._meta.fields if ((not f.primary_key) and (f.name in self.primary))] 
        where_and_clause = '' 
        for f in otpk: 
            where_and_clause = where_and_clause.join(' AND %s=%s ' % \ 
                (backend.quote_name(f.attname), 
                 str(getattr(self, f.attname)))) 
        # TODO: the last value in the above % thingy is probbably unsafe... 
         
        cursor = connection.cursor() 
         
        # First, try an UPDATE. If that doesn't update anything, do an INSERT.
        pk_val = self._get_pk_val() 
        pk_set = bool(pk_val) 
        record_exists = True 
        if pk_set: 
            # Determine whether a record with the primary key already exists. 
            cursor.execute("SELECT 1 FROM %s WHERE %s=%%s %s LIMIT 1" % \ 
                (backend.quote_name(self._meta.db_table), 
                 backend.quote_name(self._meta.pk.column), 
                 where_and_clause), [pk_val]) 
            # If it does already exist, do an UPDATE. 
            if cursor.fetchone(): 
                db_values = [f.get_db_prep_save(f.pre_save(self, False)) for f in non_pks] 
                if db_values: 
                    cursor.execute("UPDATE %s SET %s WHERE %s=%%s %s" % \ 
                        (backend.quote_name(self._meta.db_table), 
                        ','.join(['%s=%%s' % backend.quote_name(f.column) for f in non_pks]),
                        backend.quote_name(self._meta.pk.column),
                        where_and_clause),
                        db_values + [pk_val])
            else:
                record_exists = False
        if not pk_set or not record_exists:
            field_names = [backend.quote_name(f.column) for f in self._meta.fields if not isinstance(f, AutoField)]
            db_values = [f.get_db_prep_save(f.pre_save(self, True)) for f in self._meta.fields if not isinstance(f, AutoField)]
            # If the PK has been manually set, respect that.
            if pk_set:
                field_names += [f.column for f in self._meta.fields if isinstance(f, AutoField)]
                db_values += [f.get_db_prep_save(f.pre_save(self, True)) for f in self._meta.fields if isinstance(f, AutoField)]
            placeholders = ['%s'] * len(field_names)
            if self._meta.order_with_respect_to:
                field_names.append(backend.quote_name('_order'))
                # TODO: This assumes the database supports subqueries.
                placeholders.append('(SELECT COUNT(*) FROM %s WHERE %s = %%s %s)' % \
                    (backend.quote_name(self._meta.db_table), backend.quote_name(self._meta.order_with_respect_to.column), where_and_clause))
                db_values.append(getattr(self, self._meta.order_with_respect_to.attname))
            if db_values:
                cursor.execute("INSERT INTO %s (%s) VALUES (%s)" % \
                    (backend.quote_name(self._meta.db_table), ','.join(field_names),
                    ','.join(placeholders)), db_values)
            else:
                # Create a new record with defaults for everything.
                cursor.execute("INSERT INTO %s (%s) VALUES (%s)" %
                    (backend.quote_name(self._meta.db_table),
                     backend.quote_name(self._meta.pk.column),
                     backend.get_pk_default_value()))
            if self._meta.has_auto_field and not pk_set:
                setattr(self, self._meta.pk.attname, backend.get_last_insert_id(cursor, self._meta.db_table, self._meta.pk.column))
        transaction.commit_unless_managed()

        # Run any post-save hooks.
        dispatcher.send(signal=signals.post_save, sender=self.__class__, instance=self)

As some will see, this is a copy and paste of the original save function, but with additional constraints in the where clause everywhere.

Just import and inherit from this class where you have multiple primary keys and everything will work.

Also you need some sql like this: (for postgres 8.2)

ALTER TABLE app_book DROP CONSTRAINT app_book_pkey;
ALTER TABLE app_book ADD CONSTRAINT app_book_pkey PRIMARY KEY ("shelf", "level", "index")

This is a fix that works now for the impatient (like me), but I would really have it work natively, without this hackish code, so keep the good work up dcramer!

Happy programming wishes Niklas Ulvinge

[anonymous]

Thanks, exactly what i was looking for.

Chris, B.
​Phim Online

[Ravi Gidwani <Ravi.Gidwani@…>]

Replying to jacob:

In the interest of being able to use Django as an admin interface for any db schema, it really should support multiple column primary keys.

Currently, you can "fake" it by declaring one of the keys to be primary in Django and adding a unique constraint to the model. You'd have to do without the auto-generated SQL that Django gives you, but if you're smart enough to know how and why to use multiple primary keys, you can probably write the schema by hand. Still, this is less than ideal.

This is likely to be very difficult.

################################################################################################
# Classes for models that need a composite key. Django as of 0.96 does not support composite keys
# To have composite keys effect/workarround in your model follows the steps:
# Step1: Paste the below classes in your models.py (before your composite models)
# Step2: Extend you class from CompositeKeyModel
# Step3: Define 'unique_together=(("field1","field2","foreignkey__field"),) in your models Meta class
# Step4: additionally you can define unique index on the database
#
# Thats it! The save() on the instances of these models should have the composite key effect.
#
# Send feedback to: Ravi.Gidwani@gmail.com
##################################################################################################

class CompositeKeyModel(models.Model): pass

class IntermediateModelBase(ModelBase):
  def __new__(cls, name, bases, attrs):
    if CompositeKeyModel in bases:
      if bases == (CompositeKeyModel,):
        # create the class via the super method
        newclass = ModelBase.__new__(ModelBase, name, (models.Model,), attrs)
        # but then make it inherit from our model class
        newclass.__bases__ = (CompositeKeyModel,)
        return newclass
      else: raise Exception, "IntermediateModelBase does not support more than one base"
    else:
      return type.__new__(cls, name, bases, attrs)


class CompositeKeyModel(models.Model):
    __metaclass__ = IntermediateModelBase
    def save(self):
         filter = {}
         #contruct model filter on the fly for the fields that listed in the
         # 'unique_together' meta attribute
         for field_name_list in self._meta.unique_together:
             for field in field_name_list:
                 filter[field]='%s' % (self.getFieldValue(field))     

         #use the generated filter to check if the object already exist
         # if so fetch it
         fetched = self.__class__.objects.complex_filter(filter)
         #if fetched, then get its primary key and set it into the object instance
         #that is being saved.
         if(len(fetched) > 0):
             pk = self.getPrimaryKey()
             self.__setattr__(pk.name,fetched[0].__getattribute__(pk.name))
    
         #finally call the super class i.e Model save() method   
         models.Model.save(self)
         
    def getFieldValue(self,fieldName):
        separator = fieldName.find('__')
        if(separator > -1): 
            tokens = fieldName.partition('__')
            foriegnObj = getattr(self, tokens[0])
            return getattr(foriegnObj,tokens[2])
        else:
            return getattr(self, fieldName)

    def getPrimaryKey(self):                                
        for field in self._meta.fields:
            if (field.primary_key):
                return field
        
        raise Exception('No primary key')
         
#################################################################################################         

[David Cramer]

Neat hack and all, but we should stick to the topic on trac.

[Ravi Gidwani <Ravi.Gidwani@…>]

I agree, but I am sure it can be useful to someone,like me, until this bug is fixed. What say ?

[mrts]

Non-essential for 1.0.

[David Cramer]

I don't currently have time to reimplement this in qs-rf so if someone else wants to volunteer go for it.

[David Cramer]

I'm going to be working up a patch in the next few days, please see ​http://groups.google.com/group/django-developers/browse_thread/thread/4b2370a0652d9135 for discussion.

[Dan Crosta]

Is the intention to add support only for composite primary keys? What about composite, non-unique indices?

#5805 was closed as a duplicate of this ticket, but I think the intention there was different. If the fix to this ticket will not add non-unique, composite indexing to Django's model API, I suggest re-opening #5805.

[David Cramer]

I don't want this to go back and forth as a patch, simply due to how large the changeset is, but I've forked the "unofficial django git repo" and put my patch there: ​http://github.com/dcramer/django-compositepks/tree/master#

So far it allows declaration and creation of primary key models. You declare them in class Meta: primary_key = ('field', 'field', 'field').

There is no ForeignKey/Relational handlers as of right now, but the admin is mostly working.

[(none)]

Milestone post-1.0 deleted

[fongandrew]

Replying to dcrosta:
Second that. This appears to add primary-key support only. It would be nice to have non-unique composite-key indexing. It would be nice for things like the generic relations in the contenttypes app for instance.

Suggested syntax:

class MyModel(models.Model):
    field_a = models.IntegerField()
    field_b = models.IntegerField()

    class Meta:
        index_together = (('field_a', 'field_b'),)

[bohney]

Correct me if I'm wrong, but this seems to be a serious issue for those of us that are using a pre-existing database managed by a DBA that won't change the schema for the table(s) with a Composite Primary Key.

[Joey Wilhelm]

+1 on the generic multi-column field. To add to the complexity of the issue (or perhaps this is a different issue?) I have a DBA who has begun implementing multi-part foreign keys, and I would love to see these implemented as well.

Of course, the multi-part primary key is by far a higher priority for me, as I have quite a large number of tables using these, and this renders the Django admin essentially unusable.

[jfishman]

+1

[Oroku Saki]

Sorry to be a rail bird (ie, I'm not skilled enough to help), but what is the rough ETA for this solution?

[unho]

I was very excited about using Django to develop a new web app, but this particular issue makes impossible to use Django for developing it.

What we need is a way to define a primary key for a model which consist on two foreign keys, one from another model and a second from another different model than the first two I mentioned. Of course we will be referring to the first model (the one with a multi-part primary key made of two foreign keys) from a fourth model, in other words, the fourth model has a foreign key pointing to the first model.

I hope you can fix this for other projects since we won't use Django for developing our app.

[unho]

And of course it is also needed for several to several relationships which have extra attributes, because they also have compound primary keys that also are foreign keys.

[mikmani]

Lack of support for composite primary key fields is a major drawback of Django and should be dealt with as a priority. Most applications use databases with tables having composite primary keys. Such applications cannot be migrated to Django which reduces Django's usability.

[pjroberts]

I'd like to add my voice to those that want this. If you need a linking table between two tables in a many-to-many relationship, it seems absurd to have to add a redundant integer primary key to an already-unique row.

[Antti Kaihola]

Adding support for composite fields and primary keys into the Django model layer seems to be part of Google's Summer of Code 2011 (see ​http://www.djangoproject.com/weblog/2011/apr/25/gsoc/ announcement on the Django weblog).

[akseli.palen@…]

I really hope this feature will be implemented as soon as possible..

[Michal Petrucha]

I'll try to post progress updates here from time to time.

I just kicked off a thread on django-developers@ to discuss the final API (​https://groups.google.com/d/topic/django-developers/rF79c8z65cQ/discussion). I'll appreciate comments from anyone interested. (I'm posting a note in here because I realize not everyone reads the mailing list.)

[Michal Petrucha]

Okay, a progress update:
there's a branch on github: ​https://github.com/koniiiik/django/

Currently, you can define a composite field, set it as the primary key and basically that's it. No support in ForeignKey or any other relationship fields yet, that's what I'll work on starting with the next week. Model validation is missing as well, i. e. CompositeFields are not yet checked for correctness.

The admin should just work, try to explore its functionality and please let me know if you find something broken. I tried to test everything I saw but of course, I may have forgotten about a few details.

All in all, feel free to clone the repo and try it out, any feedback I get will push me forward.

[anonymous]

related field in composite

[anonymous]

I've tried using this when one of the component fields is a relationship field, but it doesn't seem to work. I might be doing it wrong. I've attached a small example that can be applied on top of koniiiik's existing tests.

[anonymous]

Ignore test-m2m for the time being, I didn't mean the FK target to have a composite PK. I'll try to get a useful test case out of my real model tomorrow.

[anonymous]

Follow up here: ​https://github.com/koniiiik/django/pull/1

[Erik Allik]

Why am I getting notifications about this issue even though I'm not in the CC? Or at least Trac only allows me to Add myself, not Remove.

[Erik Allik]

From now on I'm reporting code.djangoproject.com e-mails as spam as removing myself from CC has no effect whatsoever.

[Aymeric Augustin]

(off-topic for this ticket)

RaceCondition, given the comment you left on ticket 16763, it seems that you're aware of the situation. Whenever you contribute to a ticket, you will be notified of further updates by email.

We don't intend to change this behavior because many contributors rely on it. However, we have a manual procedure to remove people who no longer want to receive these emails. Just follow the instructions!.

And of course, don't post while logged-in, or your email will be added to the CC list again. As an alternative, I suggest posting without logging in and signing your messages.

[Erik Allik]

So, um, what is then the purpose of the CC list if you get messages anyway?
Basically, still, code.djangoproject.com is sending out spam; it's irrelevant if it's a feature for some people.

[Lau Bech Lauritzen]

We're still a bunch of hopeful people holding our breaths for this feature to enter django. What's the status? Looks like there's lots of activity on github, is the code at a state where it's reasonably stable? In which case I'd be happy to start testing on some real-life data.

[Michal Petrucha]

Well, the current status is, the soc2011/composite-fields branch is stable and I'm just keeping it up-to-date with master. There is some documentation, though I'm afraid it is not complete. Anyway, it only works with non-relationship fields.

As for relationship fields and the auxiliary_fk_fields branch, it is still only in the refactoring phase. I had to modify ForeignKey extensively and it's still not there but I'm trying to fix the bugs one by one. Once the whole test suite succeeds (except for CompositeField tests) I can resume working on composite fields themselves.

Note that this is quite a tedious process since most of the remaining failing test cases require individual handling, therefore I'll appreciate any help with this. Feel free to fork me on github and send pull requests. I don't always find enough free time to work on this myself, I've still got some exams to do and also a Bc. thesis to work on...

[Pawel Suwala <pawel.suwala@…>]

I have bugfixed and ported the CompositeKey base class posted by Ravi (4 years ago!) to work with current django version.

Here comes the snippet for people who can't wait for real-ck support in django.

################################################################################################
# Base Class for weak-emulation of composite primary key. 
# To have composite keys workaround in your model, follow these steps: 
# Step1: Paste this snippet before your models in models.py
# Step2: Extend you class from CompositeKeyModel
# Step3: Define 'unique_together=(("field1","field2","field3"),) in your models Meta class
# Step4: Create db-table by hand.
#
# That's it! The save() on the instances of these models should have the composite key effect.
#
# Send feedback to: pawel.suwala@fsfe.org
##################################################################################################

from django.db import models
from django.db.models.base import ModelBase

class CompositeKeyModel(models.Model): pass

class IntermediateModelBase(ModelBase):
  def __new__(cls, name, bases, attrs):
    if CompositeKeyModel in bases:
      if bases == (CompositeKeyModel,):
        # create the class via the super method
        newclass = ModelBase.__new__(ModelBase, name, (models.Model,), attrs)
        # but then make it inherit from our model class
        newclass.__bases__ = (CompositeKeyModel,)
        return newclass
      else: raise Exception, "IntermediateModelBase does not support more than one base"
    else:
      return type.__new__(cls, name, bases, attrs)


class CompositeKeyModel(models.Model):
    __metaclass__ = IntermediateModelBase
    def save(self, *args, **kwargs):
         filter = {}
         #contruct model filter on the fly for the fields that listed in the
         # 'unique_together' meta attribute

         if not self._meta.unique_together:
             raise ValueError('Specify Meta.unique_together to emulate Composite pk')

         for field_name_list in self._meta.unique_together:
             for field in field_name_list:
                 filter[field]='%s' % (self.get_field_value(field))

         #use the generated filter to check if the object already exist
         # if so fetch it
         fetched = self.__class__.objects.complex_filter(filter)
         #if fetched, then get its primary key and set it into the object instance
         #that is being saved.
         if(len(fetched) > 0):
             pk = self.get_primary_key()
             self.__setattr__(pk.name,fetched[0].__getattribute__(pk.name))

         #finally call the super class i.e Model save() method
         models.Model.save(self, *args, **kwargs)

    def _is_foreign_key(self, field): # there must be a better way to do this
        meta_class = getattr(field.__class__, '__metaclass__', None)
        return meta_class == ModelBase

    def get_field_value(self, fieldName):
        field_value = getattr(self, fieldName)
        if self._is_foreign_key(field_value):
            return field_value.pk
        return field_value

    def get_primary_key(self):
        for field in self._meta.fields:
            if (field.primary_key):
                return field

        raise Exception('Your model must have a dummy primary key (id)')

#################################################################################################
## example:
#class MyModel(CompositeKeyModel):
#    id                 = models.AutoField(primary_key=True) # dummy-for-django
#    user               = models.ForeignKey(User)
#    mess               = models.ForeignKey(Message)
#    atta               = models.ForeignKey(Attachment)
#
#    class Meta:
#        unique_together = (('user', 'mess', 'atta'),)
#
# -- You will need to make this table by hand. For postgresql:
#CREATE TABLE app_mymodel (
#    id integer NOT NULL, -- dummy, NO PrimaryKey CONSTRAINT
#    user_id integer NOT NULL, -- Foreign key to some table
#    mess_id integer NOT NULL, -- Foreign key to some table
#    atta_id integer NOT NULL  -- Foreign key to some table
#);

# -- create sequence for ID field (like MySQL auto-increment, but without pk)
# CREATE SEQUENCE app_mymodel_id_seq
#    START WITH 1
#    INCREMENT BY 1
#    NO MINVALUE
#    NO MAXVALUE
#    CACHE 1;
# ALTER SEQUENCE app_mymodel_id_seq OWNED BY app_mymodel.id;

# -- Add real composite Primary Key
# ALTER TABLE ONLY app_mymodel
# ADD CONSTRAINT comp_key PRIMARY KEY (user_id, mess_id, atta_id);


>>> MyModel.objects.create(user_id=1, mess_id=5, atta_id=7)
<MyModel: 1_5_7>
>>> MyModel.objects.create(user_id=1, mess_id=5, atta_id=7)
IntegrityError: duplicate key value violates unique constraint
DETAIL:  Key (user_id, mess_id, atta_id)=(1, 8383, 1321) already exists.

[thisgenericname@…]

From skimming over the above comments, it seems that one of the biggest issues is the representation of composite primary keys in URLs and in references to model.pk

How feasible would it be to have a set of model methods that:

1) Given a set of primary key fields, return a string representation of the composite key with the requirement that this mapping must be 1:1

2) Given a string representation of a primary key, return some sort of structure that identifies what fields should be at which values... Essentially the inverse of the above.

1 allows the generation of URLs, and allows things that examine model.pk to continue to function.

2 creates the framework necessary to support queries that dereference the pk, including lookups and possibly foreign keys.

The one requirement is that the composite-key-to-string bit must produce output that has only one possible interpretation... e.g does escaping of commas/slashes/whatever as needed. In the vast majority of cases using integer ids, composites can probably be in the form of 1.5 or 1-5 or such and escaping is a non-issue.

[Michal Petrucha]

Not at all, this is an issue that has been brought up years ago and it simply ignited the biggest discussions because this is one of the easiest things about this project and practically anyone can give their opinions on that. Personally I spent about a few hours thinking about the string representation, most of which was spent reading the comments on this ticket and then concluding that this is not really an issue.

More precisely, to implement a working string representation suitable for the admin, this is just about everything you need: ​https://github.com/koniiiik/django/blob/soc2011/composite-fields/django/db/models/fields/composite.py#L151 and ​https://github.com/koniiiik/django/blob/soc2011/composite-fields/django/db/models/fields/composite.py#L135

There are much more difficult aspects. The work is currently stalled on a refactor of ForeignKey and friends. I haven't had much time to work on this lately because it is the end of semester here and school's been keeping me busy but I still intend to work on this when I get a bit of free time. Unless someone wants to take over in which case I would be glad to assist in any way I can.

However, one more thing about the string representation, there is one nontrivial issue that hasn't been discussed in this ticket at all – GenericForeignKey. Using GenericForeignKey you can create SQL joins between the model containing a GFK and an arbitrary model. The value of a GFK pointing to a composite primary key should probably be a string, at least I don't have any better ideas. However, to be able to create a SQL join between the GFK and the composite field, there needs to be a way to reproduce the string encoding of the composite value in SQL. Something like

blog_posts.objects.annotate(Max('comments__date'))

should generate roughly this SQL query:

SELECT blog_posts.date, blog_posts.title, blog_posts.content, MAX(comments.date)
FROM blog_posts INNER JOIN comments ON (comments.object_id = PERFORM_ESCAPE(blog_posts.date, blog_posts.title))
WHERE comments.content_type = '''the content type of blog_posts'''
GROUP BY (blog_posts.date, blog_posts.title)

The important bit is PERFORM_ESCAPE(blog_posts.date, blog_posts.title) which should probably be a rather complex SQL expression that takes all the constituents of a composite field, turns them into strings, escapes them appropriately and generates one output string out of them. This expression will necessarily be strongly backend-dependent but the encoding should also be consistent across backends.

The bad thing about this is that we need to get this right on the first try. Once we release this feature with a certain encoding, we'll have to stick to that because of backwards compatibility, we can't require all users of this feature to migrate their data each time we change the encoding. With admin URLs this is not an issue as changing the string representation there would just result in slightly different URLs but the admin would still work just fine.

Actually, the last sentence is not entirely true with my current implementation because the admin also uses parts of the contenttypes framework to store the log of admin activity and currently I put the same value in there as the one used in the URL, which means that if I changed the representation, all of my logged admin activity would point to non-existing objects.

[jeremyt]

Where does this issue stand on the Django timeline? Have we settled on a solution (I like the approach of the composite field koniiiik has done). I would like to see a primary_key option being placed in the Meta class as well.

[Michal Petrucha]

Adding a primary_key option to Meta would be redundant, both the current API for single-column PKs and the proposed API for composite PKs use the primary_key boolean flag on a Field instance. And as far as I understand, attempts to add a new Meta option are usually turned down by the core team (similar to project-level settings). However, if you have good arguments, feel free to state them and this may still be considered.

As far as the progress goes, there is no timeline. The work is stalled once again for lack of time (school and other projects requiring immediate attention) but if the core team agrees, I'd like to take this to the final stages this summer as part of GSoC.

[jeremyt]

I think the primary_key option in the Meta makes since from the ease of use/flow prospective. I'm not saying that we should remove the option from the field either but it should only be specified in one space. Just like a CREATE TABLE statement really. If I have a primary key of a single column, I can specify it on that one column directly but if I have multiple, I place the primary key explanation as a constraint and not a column/field itself.

In the end, the model would be the same except the user doesn't have to manually declare the composite field(they still can if they want to though), but system would generate the composite field for them and make it the pk. But if the primary_key is missing everywhere, it continues to create the AutoField pk just as it does today.

Does this disagree with the current schema?

[Michal Petrucha]

Well… The Meta.primary_key option was the first proposed API and the outcome of ​this discussion was that the BDFL disliked this option and preferred an explicit CompositeField. The arguments for either option were mostly a matter of personal preference, therefore we settled on the one chosen by Jacob.

Personally, I don't really care. I can deliver the code for either one, there's not that much of a difference as far as the implementation is concerned. I'll start a thread on django-developers@ in the upcoming weeks regarding the status of this project and my plans, that could be a good opportunity for you to bring this up. Just keep in mind that this feature has a history of attracting lots of bikeshedding discussions while the more serious architectural issues are mostly being ignored. To me, this smells more like the former. So, whatever the final decision on django-developers@ is, I'll just accept it and write it that way.

[jeremyt]

Yeah.. the ideas on this have been plentiful as shown by that thread. The optional primary_key Meta option definitely goes with the DRY principle that Django is based upon but as long as the scenario is enabled, then I'm ok with spelling it out every time.

This patch plus the work I've done on multi-column joins https://code.djangoproject.com/ticket/19385 should open up some major use cases. I lend a helping hand to this issue, as well, to handle any concerns.

[Andrew J Ford]

Jeremy makes a good point about placing primary_key as a constraint within Meta. However, for the interest of time IMO, sticking with the current plan is the best path. Adding primary_key to Meta later as an afterthought (assuming the rest of the community finds a need for it) should be simple enough without breaking (future) existing code.

What is the difference between the auxillary_fk_fields and soc2011/composite-fields branches? How often is master merged? Obviously, this isn't a simple change. Maybe the community could help out a little if they had some status of what they could do to help.

[Michal Petrucha]

I apologize for the lack of timely response.

#19385 looks like it implements the last big piece that is missing from my work. I'll need to have a more thorough look at it though.

The current status of my work is kinda sad. The repo is badly out of date because it has become a real pain to keep it in sync with master. I won't get into much detail about why here, but I think it isn't worth the effort to try to sync it up. Basically, the problem is that I first implemented CompositeField (in branch soc2011/composite-fields) without ForeignKey support and then I started to refactor ForeignKey on top of that (auxiliary_fk_fields). This has become a real nightmare to maintain. I managed to get the branches through the initial Py3k changes but that required carefully merging each upstream commit one by one, after that there were numerous cleanups in the ORM which I didn't get around to merging.

However, that doesn't mean I'm giving up, I've got other plans. The idea is to do it the other way around, first refactor ForeignKey and only after it works with simple fields, port CompositeField on top of it. This should still make it possible to reuse most of the code I wrote throughout the past two years and make the changes easier to follow.

I'd like to apply for the upcoming Google Summer of Code with this to finally put this to rest, so keep an eye out on django-developers@ for an email in the near future.

[Asif Saifuddin Auvi]

these are the draft dep to implement composite primary key on django orm

​https://github.com/django/deps/blob/master/draft/0191-composite-fields.rst

and

​https://github.com/django/deps/blob/master/draft/0192-standalone-composite-fields.rst

anyone interested to completed this feature could have some look on them.

[Asif Saifuddin Auvi]

for any one interested

​https://github.com/django/django/pull/4553

[Asif Saifuddin Auvi]

As per conversation in IRC the existing Deps should also be revised based on other works and also ​https://gist.github.com/koniiiik/5408673

[Asif Saifuddin Auvi]

#16508 could be related

[Solomon Ucko]

It's been about 13 years since this issue was created and about 8 years since work on it has begun, and it still hasn't been implemented! Does the basic functionality at least work? FWIW, AFAIK the main motivation for this feature is "relationship tables" (i.e. ones that link 2 or more tables together), which is relatively common. Also, a better API might be allowing primary_key to be set on multiple fields. For my use case, I don't need to reference the set of columns directly from the database, just in the view logic. (The item specified by the URL and the current user are my two primary keys.) I could just use unique_together, but I'd rather not unnecessarily add an extra column. It's not that much of a problem, but I'd prefer not to need to.

[raydeal]

I would like to dive into this issue and see how can I help.

[raydeal]

I am digging in Dj ORM code and trying understand how FK works underneath.

[Clouds]

​https://github.com/django/django/pull/16075

[Clouds]

Hello, the main work of CompoisteField has been done and I'm now working on primary_together/PrimaryKeyConstraint.
While I've got several questions:

1. How to write test for migration? (or witch existing test as example)
   1. test if the model generating correct models.State
   2. test if from OldModel to NewModel would success migrated
2. Shall we have PrimaryKeyConstraint and CompositeField in one PR?
   1. primary (without multicolumn) first, then composite
   2. composite (cannot set as PK) first, then primary

The main blocker split it into 2 PR might be hard to add meaningful tests in first PR.

[Mariusz Felisiak]

#5929 was marked as a duplicate of this ticket.

[HAMA Barhamou]

Dear Members of Tickets #5929 and #373, I invite you to join a new discussion to clarify the differences and overlaps between these two tickets and address why #373 was marked as a duplicate

​https://forum.djangoproject.com/t/request-for-clarification-and-guidance-on-the-status-of-ticket-5929-following-closure-of-ticket-373/26075

[Csirmaz Bendegúz]

I have submitted a patch: ​https://github.com/django/django/pull/18056

[Csirmaz Bendegúz]

I have a patch but it's quite large.
I separated the tuple lookups to another PR for easier review, please check.

​https://github.com/django/django/pull/18404

[Sarah Boyce <42296566+sarahboyce@…>]

In 1eac690d:

Refs #373 -- Added tuple lookups.

[Csirmaz Bendegúz]

I separated another small refactoring for easier review.
This PR introduces the Field.is_set() function, please check.

​https://github.com/django/django/pull/18450

[Natalia Bidart]

Setting patch needs improvement following feedback in ​PR 18450

[Csirmaz Bendegúz]

Replying to Natalia Bidart:

Setting patch needs improvement following feedback in ​PR 18450

Thank you for the thorough review Natalia!
You're completely right. I adjusted my PR.

[Natalia Bidart]

I found a few more candidates for replacement in ​PR 18450, setting the flag until Ben confirms (or rejects :-)).

[Csirmaz Bendegúz]

Replying to Natalia Bidart:

I found a few more candidates for replacement in ​PR 18450, setting the flag until Ben confirms (or rejects :-)).

Thank you! I had some ​reasons for not changing those but now I actually think you're right, most of these can be refactored as well.

[GitHub <noreply@…>]

In 5865ff5:

Refs #373 -- Added Model._is_pk_set() abstraction to check if a Model's PK is set.

[Csirmaz Bendegúz]

Thanks Natalia!

Once again, I found two tiny adjustments that I separated into their own PRs for easier review, they should be a quick win:

1. Removed outdated comment from RelatedIn. (​#18564)
2. Added ColPairs.__repr__. (​#18565)

[Sarah Boyce <42296566+sarahboyce@…>]

In 347ab72:

Refs #373 -- Improved test coverage of tuple lookup checks.

This also removed unreachable checks.

[Sarah Boyce <42296566+sarahboyce@…>]

In 727587c0:

Refs #373 -- Refactored variable assignments in ColPairs and tuple lookups.

[GitHub <noreply@…>]

In 96c9907:

Refs #373 -- Removed outdated comment in RelatedIn lookup.

[Csirmaz Bendegúz]

I have made some ​improvements to the tuple lookups.

My goal is to make tuple lookups easier to use.
Once these improvements are merged, I can separate the supports_tuple_in_subquery feature into its own PR for easier review.

[Sarah Boyce <42296566+sarahboyce@…>]

In 9ca1f6ef:

Refs #373, Refs #24121 -- Added ColPairs.repr().

[Sarah Boyce <42296566+sarahboyce@…>]

In c2c7dbb:

Refs #373 -- Updated TupleIsNull lookup to check if any is NULL rather than all.

Regression in 1eac690d25dd49088256954d4046813daa37dc95.

[Sarah Boyce <42296566+sarahboyce@…>]

In 5ed72087:

Refs #373 -- Added support for using tuple lookups in filters.

[Csirmaz Bendegúz]

Thank you Sarah!

I have submitted a follow-up to add the tuple lookup validations that got removed previously, this time with extensive test coverage:
​Refs #373 -- Added additional validations to tuple lookups.

[Sarah Boyce <42296566+sarahboyce@…>]

In 97c05a6:

Refs #373 -- Added additional validations to tuple lookups.

[Csirmaz Bendegúz]

Thank you Sarah!

I have submitted the follow-up: ​Refs #373 -- Added TupleIn subqueries.
It separates the supports_tuple_in_subquery database feature from the main PR (​#18056) for easier review.

Cheers

[Csirmaz Bendegúz]

Replying to Mariusz Felisiak:

Summary: Add support for multiple-column primary keys → Add support for multi-columns fields.

It would be best to limit the scope of this ticket to the original goal of implementing composite primary keys.

We can include admin support + generic foreign key support as requested by Jacob, but we shouldn't extend the scope beyond that. Alternatively we could handle admin + generic fk support in their own follow-up tickets.

[Sarah Boyce <42296566+sarahboyce@…>]

In 03c0a3d:

Refs #373 -- Used a feature flag to disable composite subquery test on MySQL.

[Sarah Boyce <42296566+sarahboyce@…>]

In f7601ae:

Refs #373 -- Added TupleIn subqueries.

[Csirmaz Bendegúz]

Thanks Sarah!

I can't find anything else that can be merged separately.
I rebased ​Fixed #373 -- Added CompositePrimaryKey.

It looks big but there are only 442 lines of code changes in django, the rest are docs and tests.
Any review is appreciated.

[Ben Finney]

Replying to Csirmaz Bendegúz:

It would be best to limit the scope of this ticket to the original goal of implementing composite primary keys.

I agree with this. Can we please re-title this issue, back to “Add support for multiple-column primary keys”?

Csirmaz has an implementation of a composite primary key, and it correctly refers to this issue (#373) requesting composite primary key.

The implementation of an *individual field* with multiple columns, is not the same thing, and in principle is not required. (In practice it may be one feasible implementation, but it is not necessarily related to this one.) There is an existing issue (#5929) requesting multi-column fields, which I argue should be re-opened since it is a distinct issue.

I ask that the title change, made last year (see https://code.djangoproject.com/ticket/373?replyto=197#comment:162 ) should be reversed, as Csirmaz describes.

[Ben Finney]

Replying to Ben Finney:

The implementation of an *individual field* with multiple columns, is not the same thing, and in principle is not required. (In practice it may be one feasible implementation, but it is not necessarily related to this one.) There is an existing issue (#5929) requesting multi-column fields, which I argue should be re-opened since it is a distinct issue.

(Yes, I'm aware there was a discussion opened on a different platform to address this. That discussion is not open to anyone who can't use a GitHub account to authenticate, so I raise the matter here where I can participate.)

[Sarah Boyce]

We have received several approvals to ​https://github.com/django/django/pull/18056 and plan to land this soon
I would like to scope this ticket to composite/multi-column primary keys and create the following new tickets:

• ticket for composite/multi-column FK support (related existing PR: ​https://github.com/django/django/pull/18205)
• ticket for Generic FKs to support composite primary keys
• ticket for the admin to support Models with multi-column fields (esp. CompositePrimaryKeys)

Unless someone disagrees, I will rename this to Add support for multi-columns primary key fields. and create other tickets shortly

[Ben Finney]

Thank you for moving on this, Sarah.

Replying to Sarah Boyce:

I would like to scope this ticket to composite/multi-column primary keys and create the following new tickets:

• ticket for composite/multi-column FK support (related existing PR: ​https://github.com/django/django/pull/18205)
• ticket for Generic FKs to support composite primary keys
• ticket for the admin to support Models with multi-column fields (esp. CompositePrimaryKeys)

Yes, those are great.

Unless someone disagrees, I will rename this to Add support for multi-columns primary key fields. and create other tickets shortly

Slight disagreement: a composite key is not a field, it's a key (composed of multiple fields). I suggest "Add support for multi-column keys" is accurate.

[Simon Charette]

Slight disagreement: a composite key is not a field, it's a key (composed of multiple fields). I suggest "Add support for multi-column keys" is accurate.

It's more than a key in this case Ben as the proposed changes lays down the foundation to refer to the primary key as a field itself. So in this case it is a field in the Django sense even if it's not a field/column in the SQL sense.

[Csirmaz Bendegúz]

Yes, my proposal is to implement composite PKs as a virtual field. This has been discussed before I came along, many believe it's the most natural way to implement this feature in Django. I pushed against it initially, but after working on the issue for a long time I had to admit it's the best approach.

[Sarah Boyce <42296566+sarahboyce@…>]

In 978aae43:

Fixed #373 -- Added CompositePrimaryKey.

Thanks Lily Foote and Simon Charette for reviews and mentoring
this Google Summer of Code 2024 project.

Co-authored-by: Simon Charette <charette.s@…>
Co-authored-by: Lily Foote <code@…>

[Claude Paroz]

Thanks a lot to all people involved in fixing this!! I think a special price could be granted to those contributing to closing the oldest Django ticket :-)

Csirmaz, you are our hero of the day (and even more)!

[Peter Thomassen]

Great achievement indeed!

​https://github.com/django/django/blob/978aae4334fa71ba78a3e94408f0f3aebde8d07c/docs/topics/composite-primary-key.txt#L129-L132 says:

ForeignObject is an internal API. This means it is not covered by our
:ref:deprecation policy <internal-release-deprecation-policy>.

However, release notes mention various deprecation stuff around ForeignObject, such as ForeignObject.get_reverse_joining_columns() in release 5.0.

So, it seems like this note in the docs is in contradiction to what's practiced. Is this a docs bug? (Then, perhaps it should be fixed before shipping it.)

(My preference would be for ForeignObject to be public API, as it's becoming more and more important -- e.g., when using a GeneratedField with a related manager, and now with composite PK.)

[Sarah Boyce]

However, release notes mention various deprecation stuff around ForeignObject, such as ForeignObject.get_reverse_joining_columns() in release 5.0.

Sometimes we decide to follow a deprecation process for things that are not documented. It's decided case by case. However, when it is part of the public api and we have to follow the deprecation policy - I think the statement is still correct

[Csirmaz Bendegúz]

Thanks! :)
I opened 4 follow-ups:

1. #35941
2. #35953
3. #35956
4. #35957

I'm planning to work on the first two in the coming weeks.

[Sarah Boyce <42296566+sarahboyce@…>]

In 81cf690:

Refs #373 -- Fixed CompositePKChecksTests.test_composite_pk_cannot_include_generated_field() test crash on databases with no GeneratedField support.

[Sarah Boyce <42296566+sarahboyce@…>]

In 49761ac9:

Refs #373 -- Simplified DatabaseIntrospection.get_constraints() tests for composite primary keys.

[Sarah Boyce <42296566+sarahboyce@…>]

In 28f81a1:

Refs #373 -- Fixed CompositePrimaryKey tests if yaml isn't installed.

[Csirmaz Bendegúz]

#5929 was marked as a duplicate of this ticket.
I think we should re-open it and treat it as a follow-up to #373 (#5929 - Add generic CompositeFields).
I'm not planning to work on this, but maybe someone else can pick it up.