Fixed #19463 -- Add UUIDField

[mjtamlyn]

Added UUIDField. Not particularly complicated this one - no custom lookups or anything. Casts all values to uuid.UUID instances in python.

TODO:

• More form field tests
• Documentation

Squashed patch should be marked as fixing #19463.

[timgraham]

I'd alphabetize this before 'django.core'

[timgraham]

Python's
shorter: ":class:~python:uuid.UUID"

[timgraham]

.objects.create()?

[mjtamlyn]

Not much difference either way I guess. The ArrayField tests were committed with it this way. .create() doesn't really do any magic so I can change them all if you like.

[timgraham]

I think create() is cleaner, but I wouldn't change the others as part of this PR.

[mjtamlyn]

I'll leave it be for now and then add an extra commit later changing all the tests.

[timgraham]

what does cm stand for?

[mjtamlyn]

context manager?

[timgraham]

oh, I guess I'm more used to seeing as e for "exception". I see both are used in our tests; thanks.

[timgraham]

doesn't clean() return a valid you could check?

[mjtamlyn]

It returns the value, I'm not sure there's any benefit in checking it.

[timgraham]

LGTM.

[timgraham]

LGTM.

[akaariai]

I don't feel that good about adding UUID field for PostgreSQL only - UUID field isn't at all postgresql specific feature. OTOH I do understand why you want to implement it as a PostgreSQL only feature.

Also, UUID field is almost always used as a primary key, so it should be tested as a primary key. The field should get an automatic value when saving a new object into the database (the value can be generated in Python).

I recall a random discussion some time ago on IRC about problems with custom UUID field and admin - if we add UUID field, it should work automatically in admin, too. Unfortunately I don't recall details about this discussion, it was likely about using UUID as a primary key in admin. Maybe we need some simple tests with UUID as primary key in admin?

[akaariai]

I don't feel that good about adding UUID field for PostgreSQL only - UUID field isn't at all postgresql specific feature. OTOH I do understand why you want to implement it as a PostgreSQL only feature.

Also, UUID field is almost always used as a primary key, so it should be tested as a primary key. The field should get an automatic value when saving a new object into the database (the value can be generated in Python).

I recall a random discussion some time ago on IRC about problems with custom UUID field and admin - if we add UUID field, it should work automatically in admin, too. Unfortunately I don't recall details about this discussion, it was likely about using UUID as a primary key in admin. Maybe we need some simple tests with UUID as primary key in admin?

[akaariai]

The simplest solution I can think of for getting automatically a new primary key UUID value when saving a new object is addition of a new method generate_new_pk_value(self, connection) to Field. Then, if it returns something not None, the return value is used as primary key value for the model when inserting a new model. Addition of the generated UUID to the model can be done in Model._save_table(), somewhere around "if not pk_set: fields = [f for f in fields ...].

[akaariai]

The simplest solution I can think of for getting automatically a new primary key UUID value when saving a new object is addition of a new method generate_new_pk_value(self, connection) to Field. Then, if it returns something not None, the return value is used as primary key value for the model when inserting a new model. Addition of the generated UUID to the model can be done in Model._save_table(), somewhere around "if not pk_set: fields = [f for f in fields ...].

[schmitch]

I wouldn't put that in contrib, too. When you look at the code you would see, that its not "that" hard to add to other databases. We would only changed the field based on the connection. Like postgresql uses the db_type uuid, while sqlite3 uses a char field, while mysql uses uuid, too. wouldn't be too hard since the uuid getting created in python and not in the database.

[schmitch]

I wouldn't put that in contrib, too. When you look at the code you would see, that its not "that" hard to add to other databases. We would only changed the field based on the connection. Like postgresql uses the db_type uuid, while sqlite3 uses a char field, while mysql uses uuid, too. wouldn't be too hard since the uuid getting created in python and not in the database.

[schinckel]

I'd suggest an auto keyword to the __init__ for this field type.

Actually, the more I think about it, perhaps it should be even more aggressive:

def get_default(self):
    default = super(UUIDField, self).get_default()
    if not default and not self.null:
        default = uuid.uuid4()
    return default

That is, unless we are allowing NULL, always create a value.

[schinckel]

Hmm. It occurs to me that if primary_key=True, and we do this, it would look like the object already exists in the database before it actually is saved.

Although, since postgres cannot create UUIDs, that's going to present a problem anyway.

[schmitch]

Postges can create uuids, when installing the extension.
Am 19.07.2014 15:20 schrieb "Matthew Schinckel" notifications@github.com:

In django/contrib/postgres/fields/uuid_field.py:

@@ -0,0 +1,43 @@
+import uuid
+
+from django.contrib.postgres import forms
+from django.core.exceptions import ValidationError
+from django.db.models import Field, SubfieldBase
+from django.utils import six
+from django.utils.translation import ugettext_lazy as _
+
+
+class UUIDField(six.with_metaclass(SubfieldBase, Field)):

Hmm. It occurs to me that if primary_key=True, and we do this, it would
look like the object already exists in the database before it actually is
saved.

Although, since postgres cannot create UUIDs, that's going to present a
problem anyway.

—
Reply to this email directly or view it on GitHub
https://github.com/django/django/pull/2923/files#r15143651.

[mjtamlyn]

There is no justification to be this aggressive in the field definition. Making an auto-generating uuid is as simple as UUIDField(default=uuid.uuid4). This should probably be a documented pattern.

[schinckel]

Yeah, I was wound up on the fact that my similar field was not setting the value correctly on a default, but I was doing something stupid. The default= is a better pattern (and is extensible to dates, etc).

[mjtamlyn]

I don't honestly understand why this is the only field in here which needs SubfieldBase. I know I'm using to_python but there are other fields here that do so. Perhaps I need to tweak the contribute_to_class? Or is it not an issue?

[schinckel]

My understanding of SubFieldBase is that all it does is call to_python whenever any value is assigned to the field.

Personally, I wish it were on DateField and friends: being able to pass ISO8601 formatted strings and have them be date/time objects immediately makes for neater test cases.

[schinckel]

Related to comment above: if you passed a non-string value that is not a valid uuid, this would not change the value (or raise an exception), but an invalid string would raise an exception.

I think that's a little inconsistent: but I'm not sure if it's worth worrying about.

[timgraham]

Would something like if value and not isinstance(value, uuid.UUID) be better?

[mjtamlyn]

I've added some tests and notes in the docs about UUIDField(primary_key=True, default=uuid4). I think this is a good enough pattern (and also allows the user to choose a different uuid function if they wish, or none if their application generates uuids client side always. I don't think any extra machinery other than default should be necessary.

As for the admin, I found a number of closed (fixed) tickets relating to non-integer primary keys. If we find any issues with this in the admin I'll happily resolve them.

[mjtamlyn]

I've added some tests and notes in the docs about UUIDField(primary_key=True, default=uuid4). I think this is a good enough pattern (and also allows the user to choose a different uuid function if they wish, or none if their application generates uuids client side always. I don't think any extra machinery other than default should be necessary.

As for the admin, I found a number of closed (fixed) tickets relating to non-integer primary keys. If we find any issues with this in the admin I'll happily resolve them.

[mjtamlyn]

@akaariai @timgraham further review would be appreciated.

[mjtamlyn]

@akaariai @timgraham further review would be appreciated.

[claudep]

Marc, did you try to store by default in a binary column? AFAIK PostgreSQL is using a binary field internally.

[mjtamlyn]

There's a few issues with using BinaryField:

• MySQL-Python for python 3 has no support
• According to the docs they can't be filtered on, which would be a fairly significant flaw
• They data format stored would be different to that likely used in string representations (i.e. in URLs, json etc) which would be the hex format. This means a user (or Django) would have to convert that string into a uuid.UUID object and then get the binary representation. I guess we might be doing this anyway.
• Aside from space in the database, I don't see any particular benefit if doing this.

[mjtamlyn]

TBH, I haven't tried it out, it may prove to be a good idea, but I'd want some significant motivation for it.

[schmitch]

This looks really good. The only things I scared about are Forms with UUIDs, thats somewhat unnormal to enter a UUID as a user. Also the explicit need for default= is somewhat "strange".

And it's too bad that this won't go into 1.7

[schmitch]

This looks really good. The only things I scared about are Forms with UUIDs, thats somewhat unnormal to enter a UUID as a user. Also the explicit need for default= is somewhat "strange".

And it's too bad that this won't go into 1.7

[mjtamlyn]

If this is in core, should it be using internal types instead?

[timgraham]

.. versionadded:: 1.8

[timgraham]

one space between sections

[timgraham]

identifiers

[timgraham]

one space between sections

[timgraham]

.. versionadded:: 1.8

[timgraham]

([max_length=36, **options])

[mjtamlyn]

Updated fixing docs issues.

[mjtamlyn]

Updated fixing docs issues.

[schmitch]

the build is failing hard due to git errors, not sure why only some are failing..
I've run the failing builts on my box and it failed (I didn't rebased)
so for me this looks merge ready..
Maybe you should rebase manually..

[schmitch]

the build is failing hard due to git errors, not sure why only some are failing..
I've run the failing builts on my box and it failed (I didn't rebased)
so for me this looks merge ready..
Maybe you should rebase manually..

[kezabelle]

Is there a reason that I'm missing (probably) that the field's length is 36 (and each DB's backend definition is 36, with the exception of postgres, which has the native datatype)?
Removing the dashes (if a string) or calling .hex turns the value into 32 chars, which both uuid.UUID() and postgres' datatype support as a input format, and both use the dash-inclusive output format.

[schmitch]

oh that's what i've missed. thats true there is something wrong on that line

[mjtamlyn]

Good point, the 36 is with the hyphens. I'll correct that

[kezabelle]

As using UUIDs as primary keys is under test, and thus they may end up used by FKs/M2Ms, it might be prudent to see if ticket 22382 (UUIDs and prefetching) is:

• valid [I concede it might not be]
• not a problem in this case, either because of the implementation itself or because of changes elsewhere in Django.

I appreciate that's an extra burden to consider, and isn't directly related to this Field implementation, but should it be a problem, it's probably best caught early in the Field's lifecycle.

[mjtamlyn]

This is definitely a valid issue, and does apply to this situation. There is a related issue that fields using SubfieldBase do not work as you might expect in values() calls. @akaariai and I have some ideas as to how to solve this and I am planning on working on it. Combining it with this patch though is likely to be confusing.

[mjtamlyn]

Same project using UUIDField as a pk/"secondary" key/general field at https://github.com/mjtamlyn/uuid-test

It seems the correct signature for a primary key is:
id = UUIDField(primary_key=True, default=uuid.uuid4, editable=False).

[mjtamlyn]

Same project using UUIDField as a pk/"secondary" key/general field at https://github.com/mjtamlyn/uuid-test

It seems the correct signature for a primary key is:
id = UUIDField(primary_key=True, default=uuid.uuid4, editable=False).

[collinanderson]

would it make sense to lazy-load uuid as most projects won't need it? It could be done with just one more line of code in this file and in django/forms/fields.py

[mjtamlyn]

You could make the same argument for all kinds of modules throughout django - starting with the import of decimal two lines above.

Python's uuid is a single python file. The overhead to import it at server-start time is so negligible it's not worth doing anything.

[collinanderson]

It's true, but decimal and datetime are much more popular.

[mjtamlyn]

Rebased against master.

Should be considered blocked by #3047 so I can remove SubFieldBase

[mjtamlyn]

Rebased against master.

Should be considered blocked by #3047 so I can remove SubFieldBase

[timgraham]

(same pattern as above, if you change it)

[timgraham]

this tripped me up when I first read it. maybe: "Note that a callable (with the parentheses omitted) is passed to default." would be a bit clearer

[timgraham]

Please add 1.8 release notes, otherwise LGTM.

[timgraham]

Please add 1.8 release notes, otherwise LGTM.

[timgraham]

model_fields.test_uuid.TestSaveLoad.test_null_handling fails on Oracle.

[timgraham]

model_fields.test_uuid.TestSaveLoad.test_null_handling fails on Oracle.

[mjtamlyn]

I thought my last commit should fix that (ed970d4)... I'm not sure if the CI did something weird though because the list of commits on the oracle build is not the same as the list on the main build page...

[mjtamlyn]

I thought my last commit should fix that (ed970d4)... I'm not sure if the CI did something weird though because the list of commits on the oracle build is not the same as the list on the main build page...

[timgraham]

Hm, I don't have insight. Push some changes and try it again.

[timgraham]

Hm, I don't have insight. Push some changes and try it again.

[mjtamlyn]

Squashed into a single commit with release notes. IMO, this is now RFC.

[mjtamlyn]

Squashed into a single commit with release notes. IMO, this is now RFC.

[timgraham]

Please add to release notes as well.

[timgraham]

should it be [max_length=32, **options]?

[mjtamlyn]

No, the max_length argument is ignored (or forced) to be 32.

[timgraham]

no comma

[timgraham]

Ship it!

[timgraham]

Ship it!

[mjtamlyn]

Merged in ed78212

[mjtamlyn]

Merged in ed78212