+1 for the INSERT and UPDATE parts.
For the .attach() method, I wonder if it would be better to work on having
generic subquery support in the ORM instead of targeting CTEs specifically.
>From Django's perspective there isn't any big difference between:
WITH fooquery AS (SELECT * FROM foo WHERE ...)
SELECT * FROM baz JOIN fooquery ON ...
and
SELECT * FROM baz JOIN (SELECT * FROM foo WHERE ...) fooquery ON ...
The great thing about the subquery form is that it works on every database,
and to my knowledge the subquery support is surprisingly standard on all
database Django needs to care about.
If we want full support for either of the above cases, then we need some
way to tell to the Django's ORM what the fooquery (either as subquery or
with CTE) produces. For simple cases this won't be complex (say, for
.values() it's just a list of expressions), but what to do when the
subquery/CTE contains .select_related() for example? Then we'd have a
situation where the subquery produces a list of expressions but those
expressions might refer to different models in the CTE.
So, the starting point should be to allow only "simple" queries in
.attach(). This means that the attached query must be either .values()
queryset, or a queryset containing expressions from single model only (no
.extra(), .annotate() or select_related() added). In addition I think
.attach() should support subqueries instead of using only CTEs. We get
support on every backend instead of PostgreSQL only for pretty much the
same amount of effort.
- Anssi
On Saturday, March 18, 2017 at 11:28:53 AM UTC+2, Josh Smeaton wrote:
>
> Thanks for bringing this up Ashley, and for all of the detail you
> provided. I'd certainly like to see CTEs make their way into Django,
> provided we could come up with a nice enough API. From the look of it,
> you've already got something that works with an okay API so I'm hopeful.
>
> I'd be very interested in seeing your POC too if you're able to share.
>
> From looking very briefly at django-cte-trees it doesn't aim to support
> user defined CTEs for anything other than recursive queries. I'd be
> interested in seeing, as part of a DEP, how CTE inclusion in django core
> could support the cte-trees project from an API perspective.
>
> On Friday, 17 March 2017 22:28:17 UTC+11, Ashley Waite wrote:
>>
>> Hey all,
>>
>>
>> I'd like to suggest adding Common Table Expression (CTE) query generation
>> as a feature to Django.
>>
>> I've been working on a project that required manipulation of many records
>> at once, and as with many ORMs found that this wasn't an ideal use-case in
>> Django. As the rest of our code base and related projects are in Django,
>> there was a strong preference to find a way to do it and keep to the same
>> model-is-the-truth design.
>>
>> I first did this by writing some hackish functions using raw querysets
>> and generating my own CTE based queries, but it lacked ideal flexibility
>> and maintainability. So I've now written some modifications into my Django
>> to do this in a more Django-esque way and think that this functionality
>> would be beneficial within the project itself, but am unsure exactly where
>> to start the conversation about that.
>>
>>
>> *Why generate CTE based queries from querysets?*
>>
>> By allowing querysets to be attached to each other, and setting
>> appropriate WHERE clauses, arbitrary and nested SQL queries can be
>> generated. Where the results of the queries are only necessary for the
>> execution of following queries this saves a very substantial amount of time
>> and database work. Once these features exist, other functionality can also
>> transparently use these to generate more efficient queries (such as large
>> IN clauses).
>>
>> This allows several powerful use cases I think Django would benefit from:
>>
>>
>> *Large 'IN' clauses*, can be implemented as CTEs reducing expensive
>> lookups to a single CTE INNER JOIN. For sets of thousands to match from
>> tables of millions of records this can be a very substantial gain.
>>
>>
>> *Composite 'IN' conditions,* where multiple fields must match and you're
>> matching against a large set of condition rows. In my usage this was "where
>> the md5/sha hashes match one of the million md5/sha tuples in my match
>> set". This is simply a CTE JOIN with two clauses in the WHERE.
>>
>>
>> *Nested data creation*, where the parent doesn't yet exist. Django
>> doesn't currently do this as the primary keys are needed, and this makes
>> normalised data structures unappealing. Using INSERTs as CTEs that supply
>> those keys to following statements means that entire nested data structures
>> of new information can be recreated in the database at once, efficiently
>> and atomically.
>>
>>
>> *Non-uniform UPDATE*s, such that a modified set of objects can all be
>> updated with different data at the same time by utilising a CTE values
>> statement JOINed to the UPDATE statement. As there's currently no way to do
>> this kind of bulk update the alternative is to update each instance
>> individually, and this doesn't scale well.
>>
>> These could also be used with aggregations and other calculated fields to
>> create complex queries that aren't possible at the moment.
>>
>>
>> *What my PoC looks like*
>>
>> With another mildly hackish PoC that creates a VALUEs set from a
>> dict/namedtuple which can be used to provide large input data, my present
>> modified version syntax looks a bit like this (not perfect queries):
>>
>> class Hashes(models.Model):
>> md5 = models.UUIDField(verbose_name="MD5 hash (base16)", db_index=True)
>> sha2 = models.CharField(max_length=44, null=True, verbose_name="SHA256
>> hash (base64)")
>>
>> # Mock QuerySet of values
>> q_mo = Hashes.as_literal(input_hashes).values("md5", "sha2")
>> # A big IN query
>> q_in = Hashes.objects.attach(q_mo).filter(md5=q_mo.ref("md5"))
>>
>> # Matched existing values with composite 'IN' (where md5 and sha2 match, or
>> md5 matches and existing record lacks sha2)
>> q_ex =
>> Hashes.objects.attach(q_mo).filter(md5=q_mo.ref("md5")).filter(Q(sha160=q_mo.ref("sha160"))
>> | Q(sha160=None))
>>
>> # Create new records that don't exist
>> q_cr = Hashes.objects.attach(q_mo,
>> q_ex).filter(md5=q_mo.ref("md5")).exclude(md5=q_ex.ref("md5")).values("md5",
>> "sha2").as_insert()
>>
>> Returning the newly created records.
>>
>> SQL can be generated that looks something like this:
>>
>> WITH cte_1_0 (md5, sha2) AS (
>> VALUES ('00002d30243bfe9d06673765c432c2bd'::uuid,
>> 'fsA8okuCuq9KybxqcAzNdjlIyAx1QJjTPdf1ZFK/hDI='::varchar(44)),
>> ('0000f20a46e4e60338697948a0917423',
>> '6bVZgpYZtit1E32BlANWXoKnFFFDNierDSIi0SraND4=')),
>> cte_1 AS (
>> SELECT "hashes"."id", "hashes"."md5", "hashes"."sha2"
>> FROM "hashes" , "cte_1_0"
>> WHERE ("hashes"."md5" = (cte_1_0.md5) AND ("hashes"."sha2" =
>> (cte_1_0.sha2) OR "hashes"."sha2" IS NULL) ))
>> SELECT "hashes"."md5"
>> FROM "hashes" , "cte_1_0" , "cte_1"
>> WHERE ("hashes"."md5" = (cte_1_0.md5) AND NOT ("hashes"."md5" = (cte_1.md5)))
>>
>> That is:
>>
>> - A qs.as_insert() and qs.as_update() on queryset to create *lazy*
>> insert and update queries.
>> - A qs.attach() that allows querysets to be attached to other
>> querysets, and will generate them as CTE statements.
>> - A qs.ref() that returns an expression that when the query is
>> compiled will be a field reference on the CTE that represents that
>> queryset.
>> - Additional compilers on the QuerySet subclasses that these return
>> (so no changes to base compilers meaning no functionality impact to
>> existing usage)
>> - Generation of WITH clauses for attached querysets, and RETURN
>> clauses for lazy UPDATE and INSERT querysets with fields requested (via
>> values() in this case)
>>
>> As these can be attached to querysets that are attached to querysets,
>> that are... etc, many statements can be chained allowing substantial
>> changes to be performed without needing Django to have to receive, process,
>> and resend at every step.
>>
>> I've had a read through the enhancement proposal docs etc, and I'm
>> willing to do what's needed to make this functionality solid, and put forth
>> a proposal to add it. But am first seeking feedback on it, and whether this
>> is a feature that will be considered.
>>
>>
>> Thanks,
>> - Ashley
>>
>
--
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 https://groups.google.com/group/django-developers.
To view this discussion on the web visit
https://groups.google.com/d/msgid/django-developers/bf796fae-6f33-4293-a4d8-70f82fb40414%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
