@vlado @simi looks like you're doing the same work at the same time. Ref: #37973

Maybe you want to compare notes as well?

@kaspth not sure how could I miss that. I have pushed also fix for update_all/delete_all in my branch and it is missing recursive for now (to keep the PR simple).

@vlado would you mind if I port some of your changes into my PR, mark you as an author and continue in there? You have really nice documentation and also I can port some tests.

Hi @simi and @kaspth thanks for the feedback. What a coincidence that we worked on same thing same time :)

I forgot to mention in PR description that I already added support for recursive expressions Post.with(:recursive, cool_posts: "union to get cool posts..."). It follows Arel API for recursive calls. See more here: https://github.com/rails/rails/pull/37944/files#diff-25e9a621b650cd62a407c4ce7a117b20R77-R86.

I agree that we co-author this but I think it is easier to finish it here. From what I see only thing missing is update_all / delete_all support?

@kaspth What do you think?

@vlado the recursive api looks a little weird for me. I don't think there's any other AR relation method which uses "argument as a modifier of behaviour". Therefore I have decided in my PR to not introduce it for now to not block basic implementation and open another PR later once the first one get merged.

Comparing our code changes I have spotted those:

1. I have separated with logic to separate class (ActiveRecord::With). Would you mind to port them as well if you like that?

2. In build_arel method you have removed support for Array and for Hash your code is different. Can you explain on those changes? My idea was to not change logic/API to get 100% compatible with original ctes_in_my_pg gem to get easy transition.

Finally if you can prepare failing specs for delete_all and update_all I can submit PR to your branch to support those relation calls as well.

Hi @simi, thanks for the constructive feedback. Replies are below.

@vlado the recursive api looks a little weird for me. I don't think there's any other AR relation method which uses "argument as a modifier of behaviour". Therefore I have decided in my PR to not introduce it for now to not block basic implementation and open another PR later once the first one get merged.

Hm, you could be right here. I also can not think of any AR relation method at this moment. I initially started with with.recursive(...) but when I checked how Arel works I saw that .with(:recursive, ...) is used and I liked it cause it is much simpler to implement. I was also thinking about .with_recursive.

Comparing our code changes I have spotted those:

1. I have separated with logic to separate class (ActiveRecord::With). Would you mind to port them as well if you like that?

Looks like a good idea 👍

2. In build_arel method you have removed support for Array and for Hash your code is different. Can you explain on those changes? My idea was to not change logic/API to get 100% compatible with original ctes_in_my_pg gem to get easy transition.

Arrays are still supported but only if all items are Arel::Nodes::As. I didn't add documentation for this cause I didn't see options where you can pass Arel nodes documented anywhere (probably because Arel API is considered private).

Regarding Hash changes, I think there was an issue with select = Arel::Nodes::SqlLiteral.new("(#{expression.to_sql})") line in one of the tests in SQLite (I can check it later). Also, I tried to add meaningful error messages if unsupported arguments are passed.

I never used https://github.com/DavyJonesLocker/postgres_ext or https://github.com/kmurph73/ctes_in_my_pg in any of my apps (i was using my own patch) but I used them as inspiration when starting this PR.
That being said, I think I haven't break any compatibility with them in case of the upgrade, and even if we break some it should be super easy to upgrade. I think we should prioritise to make this best it can be.

Finally if you can prepare failing specs for delete_all and update_all I can submit PR to your branch to support those relation calls as well.

Sure, I'll add failing tests.

Post.with(
  posts_with_comments: Post.where("comments_count > ?", 0),
  posts_with_tags: Post.where("tags_count > ?", 0)
)

I'm not too keen on this API. Seems verbose, yet still fairly low level. I haven't used CTEs before but just realized a case where I could, what's the use cases you've needed them for? And how do you act on the values? I'm guessing posts_with_comments/posts_with_tags are available to use in a custom SQL string?

What I'm saying is since Rails is an extracted, not built, framework, I'd prefer to see full production samples (with renames or some business logic stripped if needed). It seems like with is only half the equation, but there's also selecting on the resulting CTE. Plus what people can pass to with also includes a UNION ALL, I'd be curious to see those cases too. All in the name of figuring out how to cut and where to scope the API. Active Record isn't necessarily just Arel but nicer, it's free to be more geared. 😊

@kaspth CTEs are super useful since they work as a "temporary table". Our codebase is full of CTEs and this will make it much nicer, since we usually need to fallback to just executing SQL string written in heredoc. Also queries using CTE are often more readable. This will allow to compose those queries based on AR scopes instead of messing with SQL strings (we usually interpolate SQL string with relation.to_sql).

Yes, I agree all examples in this (and my PR) are really simple and not real reason to use CTE. I think there are articles around describing CTE and how/when to use them. On the other side Rails active record tests are full of examples not really making sense, but doing great job testing the feature.

I understand this not really well known feature of SQL (mainly for beginners). I have learned about that just few months ago. Introducing this feature with nice and friendly guide explaning purpose of this Active Record API will be great service to help people understanding CTEs.

What I'm saying is since Rails is an extracted, not built, framework, I'd prefer to see full production samples (with renames or some business logic stripped if needed).

Just one quick example I do remember is we need to group and filter data and then join them with belongs_to relation. I'll try to prepare some examples with current code and how it will work with current pull request to show you some real examples extracted from real application.

ℹ️ This is also super useful in combination with recently introduced bulk upserts in combination with RETURNING (which did not make it yet to Rails codebase).

Great questions @kaspth and you are guessing right :)

I'll try to explain with one example

Example

At BetterDoc we have a 60+ lines SQL query that we use to dynamically calculate clinic rank based on some params (medical problem, geo location and radius, ...).
Keep in mind that this is a super simplified version that should be enough to explain the use case.

Domain

We have 2 models in our domain. Clinic with name only, and Case Number which stores number of operations (count) clinic did for specific medical topic (code).

class Clinic
  has_many :case_numbers
end

class CaseNumber
  belongs_to :clinic
end

SQL Query

Here is a full SQL that is explained below.

WITH relevant_clinics AS (
    SELECT clinics.id, clinics.name, SUM(case_numbers.count) AS cases_count 
    FROM clinics INNER JOIN case_numbers ON case_numbers.clinic_id = clinics.id
    WHERE case_numbers.code ILIKE 'F4%'
    GROUP BY clinics.id 
),
stats AS (
    SELECT
        PERCENTILE_CONT(0.25) WITHIN GROUP (ORDER BY cases_count DESC) AS quartile_2,
        PERCENTILE_CONT(0.50) WITHIN GROUP (ORDER BY cases_count DESC) AS quartile_1
    FROM relevant_clinics
)

SELECT name, cases_count,
    CASE 
        WHEN cases_count > (SELECT quartile_2 FROM stats) THEN 'High'
        WHEN cases_count > (SELECT quartile_1 FROM stats) THEN 'Average'
        ELSE 'Low' 
    END AS rank
FROM relevant_clinics
ORDER BY cases_count DESC

First step is to get clinics that match provided medical topic and store them as temporary result set (relevant_clinics). Then we reference them to get some statistics from them and save those as another temporary results set (stats). We then fire main query which references both of them to get final results with rank and other stuff we need.

How to organise this in Rails

Since ActiveRecord does not have support for CTE we started thinking how to organise this in our app.

We started by putting SQL into heredoc and executing it with find_by_sql and count_by_sql. This could work for some report or similar but it didn't work for us cause we needed ActiveRecord::Relation flexibility (chainability, lazy load, ...).

Next try was with Arel.with and it was a step forward but we were still missing ActiveRecord::Relation goodies. Clinic.where(Clinic.arel_table[:id].in(arel_that_returns_clinic_ids)) "hack" was not an option cause we needed to keep order and other stuff from original query.

In the end I wrote a monkey patch that adds .with to ActiveRecord::Relation and we were able to write our query like this.

class Clinic < ApplicationRecord
  has_many :case_numbers

  scope :matches_medical_topic, -> (q) { joins(:case_numbers).where("case_numbers.code ILIKE ? "#{q}%") }

  def self.search(q)
    relevant_clinics = matches_medical_topic(q)
      .select("cllinics.id, clinics.name, sum(case_numbers.count) AS cases_count
      .group("clinics.id")

    stats = select(
        "PERCENTILE_CONT(0.25) WITHIN GROUP (ORDER BY cases_count DESC) AS quartile_2",
        "PERCENTILE_CONT(0.50) WITHIN GROUP (ORDER BY cases_count DESC) AS quartile_1"
      )
      .from("relevant_clinics")

    with(relevant_clinics: relevant_clinics, stats: stats)
      .from("relevant_clinics AS clinics")
      .order("cases_count DESC")
  end

  def self.with_rank
    select(:id, :name, :cases_count)
      .select("CASE WHEN cases_count > (SELECT quartile_2 FROM stats) THEN 'High' WHEN cases_count > (SELECT quartile_1 FROM stats) THEN 'Average' ELSE 'Low' END as rank")
  end
end

and use it like:

Clinic.search("f4").with_rank # => ActiveRecord::Relation
Clinic.search("f4").count # => Integer

Since there are gems to do same thing I assumed that other people would like to see this in Rails so I started this PR.

Conclusion

You can see from my example that I used from("relevant_clinics AS clinics") to be able to chain it with other relations. Alternative is to use join joins("JOIN relevant_clinics ON relevant_clinics.id = clinics.id").

I absolutely agree that this is a bit too verbose and low level but I haven't been able to figure out better API yet. I'm eager to hear your ideas and to make it more "geared" :)

Maybe the way to go could would be to have something similar to associations and/or scopes that once defined will abstract everything for the user. Similar to the way joins(:association_name)works. Option to go low level (current implementation) could also stay.

class Post
  has_many :case_numbers

  common_table :relevant_clinics, -> (q) { join(:case_numbers).where("code = ?", q) }
  common_table :stats, -> { select(...).from(:relevant_clinics) }
end

Post.with(:relevant_clinics, :stats)

I see now that I forgot about UNION example. You can find one in tests here https://github.com/rails/rails/pull/37944/files#diff-25e9a621b650cd62a407c4ce7a117b20R77-R86

Not a real life example but it is getting late here in Croatia :)

@vlado thanks for very detailed info. Our usecase and story is really similar, but in the end I have used https://github.com/kmurph73/ctes_in_my_pg as a Relation#with implementation.

@kaspth my idea was to introduce this with logic into AR step by step. Start with simple with, then continue with recursive if it will be needed. We can also later see how with is doing and if there are some new patterns like that "scope-like" common_table idea mentioned in @vlado 's comment.

I think demand for WITH support is not neglible. You can google for cte rails active record site:stackoverflow.com (sorry, I have no idea how to share search result link). There are also some blog post mentioning how to integrate CTEs into ActiveRecord.

Few examples including which strategy is used to integrate CTE to ActiveRecord (just selected from first search page):

• https://sonnym.github.io/2017/06/05/common-table-expressions-in-activerecord-a-case-study-of-quantiles/ (custom SQL in select and from + arel based solution)
• https://typefast.blog/common-table-expressions-within-rails/ (find_by_sql for ActiveAdmin)
• https://www.scimedsolutions.com/articles/75-arel-part-ii-common-table-expressions (arel based solution)
• https://hashrocket.com/blog/posts/recursive-sql-in-activerecord (mixing to_sql on scope and AR query methods, using recursive form of CTE)
• https://hashrocket.com/blog/posts/writable-common-table-expressions (using connection.select_values, explaining how to chain multiple insert/update/delete queries together, this PR is not really related to this use case; it will be really nice to support but it will need some time to bring some nice API for this; we use this pattern often)
• https://blog.castle.co/tech/10-06-2016/complex-sql-in-rails (using sql and connection.select_all, also using views to hide CTEs into DB - this approach is not usable if you need to build the WITH statement on the fly).

PS: I was thinking about gemifying this as well in some general way (not supporting only PostgreSQL), but monkeypatching of private methods from ActiveRecord and Arel would be needed. That was the main motivation to send my pull request to integrate this into core.

@vlado thanks for very detailed info. Our usecase and story is really similar, but in the end I have used https://github.com/kmurph73/ctes_in_my_pg as a Relation#with implementation.

I found out about ctes_in_my_pg late, postgres_ext was not working with Rails 6 so I did my own patch.

I think demand for WITH support is not neglible.

Totally agree. We are using CTEs with Rails for more then 4 years now and it is a pain from the beginning. I was also giving talks about this topic (organising complex SQL queries in Rails) few times and from the feedback regarding CTE I was actually surprised that this was not implemented yet and decided to do it properly this time :)

@kaspth my idea was to introduce this with logic into AR step by step. Start with simple with, then continue with recursive if it will be needed. We can also later see how with is doing and if there are some new patterns like that "scope-like" common_table idea mentioned in @vlado 's comment.

That's pretty much the opposite of what I'm curious to explore here or how I'd proceed. I'd rather we have as many real world cases to base this off of and extract the commonalities. Then if that looks like something that's worth it, we can consider adding that and some lower level constructs. As it is I don't think there's enough meat here to add with and I'm thinking of bowing out and passing to someone else on the team who feels more strongly about this.

For instance the common_table abstraction doesn't have much sway with me currently, from what I've gathered CTEs are for one off queries, so I'd personally rather put all those parts together.

I think demand for WITH support is not neglible. You can google for cte rails active record site:stackoverflow.com (sorry, I have no idea how to share search result link). There are also some blog post mentioning how to integrate CTEs into ActiveRecord.

Sure, unfortunately that doesn't necessarily mean that we want to maintain something in Active Record directly. Then we'd be looking at APIs for gems instead, of which the current proposed with seems to fit with better.

@vlado I'm curious about your usage of from in your examples, can you expand on that? Particularly where you alias the relevant_clinics into clinics?

Another thing that's tough for me, although I like the with_rank split, is that cases_count isn't defined in there. So calling with_rank is dependent on the other method first.

@kaspth just try to take a look at this feature from different perspective.WITH is just another SQL clause, same as WHERE, FROM, JOIN or ORDER for example. WITH was introduced in SQL-99 standard and is widely supported by all "major" relation databases.

Common table expressions are supported by Teradata, DB2, Firebird,[8] Microsoft SQL Server, Oracle (with recursion since 11g release 2), PostgreSQL (since 8.4), MariaDB (since 10.2), MySQL (since 8.0), SQLite (since 3.8.3), HyperSQL and H2 (experimental).[9] Oracle calls CTEs "subquery factoring".[10]

https://en.wikipedia.org/wiki/Hierarchical_and_recursive_queries_in_SQL#Common_table_expression

My motivation to support WITH in ActiveRecord is to make it possible to compose queries using WITH (for any reason) via nice ActiveRecord chainable API. The same you can do with custom joins, selects, wheres and orders today.

Sometimes you can avoid SQL oriented approach and just ask ActiveRecord for data like Account.all.eager_load(:users) and let the ActiveRecord do the best to build the query for you. Sometimes you need to use more SQL oriented approach and build chain of few custom selects, joins, wheres, ... and get the desired result. IMHO both approaches are totally valid. ActiveRecord excels in both of them already. Introducing WITH support will make the second approach even more flexible.

Another thing that's tough for me, although I like the with_rank split, is that cases_count isn't defined in there. So calling with_rank is dependent on the other method first.

Yes, but I think this is something you need to handle carefully already if you use more SQL oriented approach. For example if you select or order by relation column, you need to ensure related table is properly joined.

Sure, unfortunately that doesn't necessarily mean that we want to maintain something in Active Record directly. Then we'd be looking at APIs for gems instead, of which the current proposed with seems to fit with better.

Supporting this in gem is really hard since ActiveRecord internals are changing often and this needs to monkey patch various lower level APIs including AREL ones. There's solid support in AREL already for WITH. I'll extract my AREL change to make this work on update_all/delete_all as well to separated pull request (not introducing Relation#with, but just fixing this in AREL for UpdateManager and DeleteManager). Having this change merged will make it at least a little more easier to maintain in gemified form and AREL can benefit from that as well.

For instance the common_table abstraction doesn't have much sway with me currently, from what I've gathered CTEs are for one off queries, so I'd personally rather put all those parts together.

As it is I don't think there's enough meat here to add with and I'm thinking of bowing out and passing to someone else on the team who feels more strongly about this.

CTEs are not definitely useful just for one off queries. We use them for regular queries for various reasons already mentioned before. If you would like to not continue in this review I would like to thank you anyway for your time spent already on this. I can try to bring some attention here to get someone another to review if this feature is not actually finally rejected by that decision.

PS: Originally I have forgotten to mention that it is probably possible (I'm not sure if that will be possible for all adapters) to fix some problems using WITH internally. For example my work on #35638 lacks support for "through" associations. Using WITH can fix that by creating associations in WITH part, return associated ids and make entries in "linking" table in second using ids returned from CTE. That's also one of the usecases we use. Pipe queries together and use output of one as a input to another one without need to send data from DB to app and back.

@vlado I'm curious about your usage of from in your examples, can you expand on that Particularly where you alias the relevant_clinics into clinics?

@kaspth I'm aliasing relevant_clinics as clinics here to be able to easily chain and call other query methods later. For example to do something like:

Clinic.search("f4").with_rank.where(id: [1, 2, 3])
# WITH relevant_clinics AS (...) SELECT ... FROM relevant_clinics clinics WHERE clinics.id IN (1,2,3)

As mentioned you could also use join joins("JOIN relevant_clinics ON relevant_clinics.id = clinics.id")

Another thing that's tough for me, although I like the with_rank split, is that cases_count isn't defined in there. So calling with_rank is dependent on the other method first.

Agreed, I also did not liked it but it was good enough :)

Really appreciate your participation on this PR and would not want you to quit. You asked great questions and I like your thinking.

I would say that biggest challenge (at least it was for us) is that you have expression(s) that builds your temporary table(s) and then you have multiple ways you want to use them (select without rank, select with rank, aggregate, call some additional where, ...) which makes it hard to put all this things together in one abstraction. That is way I proposed the common_table abstraction which is responsible just for the first part (half the equation as you said) and then you are responsible to do something with it. Similar how joins works CaseNumber.joins(:clinic).where("clinics.name ILIKE ?", "%query%"). Not saying this is perfect solution but I'm more then happy to explore this further.

That being said I must also agree with @simi that just adding with method to Relation is already a huge improvement and it makes using CTE's with Rails 100 times better :)
Also, even when we find a proper abstraction and API for common cases with should keep current functionality cause there will always be a case when you need to go low level to build your query.

PS. I'm sorry for a late reply. Was more focused on family stuff for holidays ☃️

Could you rebase? There's lots of unrelated CI failures because of the minitest update. I'll leave a bit of time for other maintainers to chime in if they want, but I don't see any objections to merging.

Also the CI failures look legit now.

Thank you @vlado. Feel free to open a second PR for recursive support.

Thank you @vlado. Feel free to open a second PR for recursive support.

Great, thanks @byroot!

Just an unsolicited comment on this thread for posterity on the syntax and need for recursive CTEs.

1. I have had a need for recursion somewhere in most non-trivial companies I have ever done as we have ended up with something that is inherently nested. Ex. permssions groups where a group can be a group member, org charts, directed acyclic graphs representing business rules, etc. It is very needed.
2. with.recursive feels very much like what I would expect. I think it is very similar conceptually to where.not.