On the core ORM side we need to make
.exclude(LessThan(F('friends__age'), 30)) do a subquery. This way
.exclude(friends__age__lt=30) does the same thing as the expression
version. This isn't that easy to do. If we just use
resolve_expression, then the friends relation will generate a join,
and then as second step do a negated filter on the joined value.
Instead we want to detect that the LessThan expression needs to be
pushed in to a subquery.
So, we need to solve:
A) A way to ask an expression if it is referencing a multijoin
(possible approach is to just have a method
"refs_multi_valued_relation(query)")
B) When the ORM sees an expression that is reffing a multijoin in an
exclude filter, then we need to push the expression in to a subquery.
A) requires some new work. This shouldn't be that hard to implement,
we just recursively ask subexpressions if they reference a multijoin.
Something like https://github.com/django/django/pull/4385 will make B)
much easier to implement.
I've been working on making Q-objects responsible for resolving
themselves. See https://github.com/django/django/pull/4801. This
should solve 3).
We don't seem to be missing any major parts. I (or some volunteer)
just need to finish the PRs, and then we should be really close to
full support for expressions in filter.
Josh: do you think we could get Lookup as expressions in to 1.10
instead of 1.11?
- Anssi
On Wed, Sep 30, 2015 at 3:46 AM, Josh Smeaton <josh.smea...@gmail.com> wrote:
> 1. Lookups should become Expressions, just as Transforms have become
> Expressions. This will let us process Lookup arguments as Expressions all
> the way the way through. I think this should be a major goal for version
> 1.11.
>
> 2. Chaining transforms is now possible since they are just Func expressions.
> Func(Func(Func('field_name'))) is no issue.
>
> 3. Sounds like an OK idea, but I haven't looked into the details enough to
> really comment. I do think we should create the correct form as early as
> possible (parsing into a chain of Lookup/Transform expressions) so we don't
> have to do parsing in multiple places. The entry points to .filter() and
> .exclude(), or their direct counterparts in sql.query sound ideal. Anssi has
> mentioned elsewhere that WhereNode's should only contain fully resolved
> expressions, so resolving will need to be done directly after parsing (or
> during).
>
> Part 1 above can be started now if you have the time or interest. We can
> nail down the particulars of part 3 while we're solving part 1. Part 1 may
> drive some of part 3.
>
> Cheers
>
>
> On Wednesday, 30 September 2015 04:49:54 UTC+10, Alexey Zankevich wrote:
>>
>> Here is a list of issues to solve to support explicit transforms and
>> lookups by filter (and exclude) methods.
>>
>> 1. Make Lookup.__init__ signature to support initialization with F objects
>> or string path (e.g. GreaterThan(F('user__id'), 10) or
>> GreaterThan('user__id', 10)), not sure it's possible to use simultaneously
>> with the current approach with lhs, rhs initialization (even with moving it
>> to a separate class method, e.g Lookup.build(lhs, rhs)), so I assume
>> creating so-called util classes which will delegate SQL-related
>> functionality to existing Lookup classes.
>>
>> 2. Chain transforms by passing them as argument:
>>
>> Lower(Unaccent(F('user__name)))
>>
>> 3. Decide if Q objects shall support explicit lookups/transforms as
>> argument as well - it's a kind of logical step, as without Q objects it will
>> not be possible to perform complicated conditions (AND, OR, NOT).
>> In that case lookup/transform parsing should be moved from QuerySet object
>> to Q object - filter will take already parsed lookup tree.
>> Example:
>>
>> Q(user__name__lower__unaccent__icontains='Bob') will internally parse it
>> and build next structure:
>>
>> Q(Icontains(Lower(Unaccent(F('user__name')))), 'Bob')
>>
>>
>> On Sunday, August 16, 2015 at 4:18:26 PM UTC+3, Alexey Zankevich wrote:
>>>
>>> Hi all,
>>>
>>> This topic is related to the current ORM query syntax with underscores.
>>> There are lots of arguing related to it, anyway it has pros and cons.
>>>
>>> Let's take a concrete example of querying a model:
>>>
>>> >>>
>>> >>> GameSession.objects.filter(user__profile__last_login_date__gte=yesterday)
>>>
>>>
>>> Pros:
>>>
>>> 1. The syntax is easy to understand
>>> 2. Can be extended with custom transforms and lookups
>>>
>>> However, there are several cons:
>>>
>>> 1. Long strings is hard to read, especially if we have fields with
>>> underscores.
>>> It's really easy to make a mistake by missing one:
>>>
>>> >>>
>>> >>> GameSession.objects.filter(user_profile__last_login_date__gte=yesterday)
>>>
>>> Not easy to catch missing underscore between user and profile, is it?
>>> Even
>>> though, it's not easy to say whether it should be "user_profile"
>>> attribute or
>>> user.profile foreign key.
>>>
>>> 2. Query strings can't be reused, thus the approach violates DRY
>>> principle.
>>> For example, we need to order results by last_login_date:
>>>
>>> >>>
>>> >>> GameSession.objects.filter(user__profile__last_login_date__gte=yesterday)
>>> >>> \
>>> .order_by('user__profile__last_login_date')
>>>
>>> We can't keep user__profile_login_date as a variable as in the first part
>>> of the
>>> expression we use a keyword argument, meanwhile in the second part - just
>>> a
>>> string. And thus we just have to type query path twice.
>>>
>>> 3. Lookup names not natural to Python language and require to be
>>> remembered or
>>> looked up in documentation. For example, "__gte" or "__lte" lookups tend
>>> to be
>>> confused with "ge" and "le" due to similarity to methods "__ge__" and
>>> "__le__".
>>>
>>> 4. Lookup keywords limited to a single argument only, very inconvenient
>>> when
>>> necessary to filter objects by range.
>>>
>>> I was thinking a lot trying to solve those issues, keeping in mind Django
>>> approaches. Finally I came up with solution to extend Q objects with dot
>>> expression syntax:
>>>
>>> >>> GameSession.objecs.filter(Q.user.profile.last_login_date >=
>>> >>> yesterday)
>>>
>>> Q is a factory instance for old-style Q objects. Accessing attribute by
>>> dot
>>> returns a child factory, calling factory will instantiate old-style Q
>>> object.
>>>
>>> >>> Q
>>> <QFactory object at 0x7f407298ee10>
>>>
>>> >>> Q.user.profile
>>> <QFactory object at 0x7f40765da310>
>>>
>>> >>> Q(user__name='Bob')
>>> <Q: (AND: ('user__name', 'Bob'))>
>>>
>>> It overrides operators, so comparing factory with value returns a related
>>> Q
>>> object:
>>>
>>> >>> Q.user.name == 'Bob'
>>> <Q: (AND: ('user__name', 'Bob'))>
>>>
>>> Factory has several helper functions for lookups which aren't related to
>>> any
>>> Python operators directly:
>>>
>>> >>> Q.user.name.icontains('Bob')
>>> <Q: (AND: ('user__name__icontains', 'Bob'))>
>>>
>>> And helper to get query path as string, which requred by order_by or
>>> select_related queryset methods:
>>>
>>> >>> Q.user.profile.last_login_date.get_path()
>>> 'user__profile__last_login_date'
>>>
>>> You can check implementation and more examples here
>>> https://github.com/Nepherhotep/django-orm-sugar
>>>
>>> How it solves issues:
>>>
>>> #1. Dots hard to confuse with underscores
>>> #2. Query paths can be reused:
>>>
>>> >>> factory = Q.user.profile.last_login_date
>>> >>> query = GameSession.objects.filter(factory >= yesterday)
>>> >>> query = query.order_by(factory.get_path())
>>>
>>> #3. Not neccessary to remember most of lookup names and use comparison
>>> operators
>>> instead.
>>> #4. Possible to use multiple keyword arguments:
>>>
>>> >>> Q.user.profile.last_login_date.in_range(from_date, to_date)
>>> <Q: (AND: ('user__profile__last_login_date__lte', from_date),
>>> ('user__profile__last_login_date__gte', to_date))>
>>>
>>>
>>> This approach looked the best for me due to several reasons:
>>>
>>> 1. It's explicit - it doesn't do anything but generating appropriate Q
>>> object.
>>> The result of comparison can be saved as Q object variable.
>>>
>>> 2. It's short - variants with using model for that will look much longer,
>>> when
>>> joining two or more filters:
>>>
>>> >>> GameSession.objects.user.profile_last_login_date >= yesterday #
>>> >>> awkward
>>>
>>> 3. Implementation will not require to change querset manager or model
>>> classes
>>>
>>> 4. Will still allow to use filters and Q class in the old way:
>>>
>>> >>> q = Q(user__profile__last_login_date__gte=yesterday)
>>>
>>> or
>>>
>>> >>>
>>> >>> GameSession.objects.filter(user__profile__last_login_date__gte=yesterday)
>>>
>>> I'd like to make it as a part of Django ORM syntax and it will not be
>>> hard to
>>> do, especially taking into account the library is already done and
>>> working.
>>> Anyway, I need your thought about the idea in general, as well as about
>>> particular things like chosen method names - "get_path", "in_range" and
>>> etc.
>>> As next step I can create a ticket in the issue tracker, or prepare DEP
>>> first.
>>> In latter case I need to find a shepherd to work with.
>>>
>>> Best regards,
>>> Alexey
>
> --
> 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 django-developers+unsubscr...@googlegroups.com.
> To post to this group, send email to django-developers@googlegroups.com.
> 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/3bad8371-f9b4-47ff-a681-0108b320e9b5%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 django-developers+unsubscr...@googlegroups.com.
To post to this group, send email to django-developers@googlegroups.com.
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/CALMtK1H0M-Zzzwtkb0KY-HXOFsq%3DyNE-APZgeBXL54DW_XKD%3DQ%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.
