Add hash return type for relation groupchain size

[marknuzz]

Motivation

Model.where(...).group(...).size returns a hash, not a scalar number.

Implementation

This is a special case where ActiveRecord::Calculations doesn't handle the size method.

https://github.com/rails/rails/blob/6f57590388ca38ed2b83bc1207a8be13a9ba2aef/activerecord/lib/active_record/relation.rb#L273-L280

# activerecord-7.1.4/lib/active_record/relation.rb
def size
  if loaded?
    records.length
  else
    count(:all)
  end
end

Model.group('created_at').count(:all) == Model.group('created_at').size
=> true

We can handle this by adding a special case for the size method as we can't rely on the Calculations class to provide the name of the method.

Tests

Yes

[amomchilov]

Is there any reason to not add :size into the CALCULATION_METHODS array itself?

[marknuzz]

Hmm I'm actually not sure how to categorize this. These arrays are mostly organized based on the class that they were defined in, except the methods defined on ActiveRecord::Relation, where we have FIND_OR_CREATE_METHODS, BUILDER_METHODS, and TO_ARRAY_METHODS. But :size (like :to_a) is not something you can call directly on the model's class, so we can't have it in CommonRelationMethods or GeneratedRelationMethods. Unlike :to_a, :size's return type is different depending on context, so it didn't seem to fit into any of the above arrays.

But now that you bring it up, this seems like one of those cases where the method name doesn't fit nicely into into an array and the method should be defined outside of the scope of the case statement looped over an array.

And when looking into how the .to_a worked, I'm now finding that the GeneratedRelationMethods#where chain typing is actually incorrect. If you give an argument like User.where(foo: 1) you do not in fact get to use the methods on the PrivateAssociationRelationWhereChain returned. Additionally, you can't call :to_a on an object that is a where chain (User.where.not.to_a), unlike what the generated rbi suggests.

I think the best thing to do here might be to take a step back and make sure this structure is actually correct, and to figure out a way to automatically test this so that this can't happen. This PR might be a symptom of a larger problem and could complicate things further, and the .to_a additions were a reasonably good example of a case where this happened due to the assumption that the existing behavior was correct, when it wasn't. Should we close this for now or what do you think?

And now I can fully appreciate how difficult the job is of the Tapioca team!

[amomchilov]

Hey Mark!

where the method name doesn't fit nicely into into an array and the method should be defined outside of the scope of the case statement looped over an array.

That seems reasonable to me.

If you give an argument like User.where(foo: 1) you do not in fact get to use the methods on the PrivateAssociationRelationWhereChain returned.

I can't test this right now, could you explain the issue?

And now I can fully appreciate how difficult the job is of the Tapioca team!

Haha thanks! Trying to tame Ruby meta-programming with types is a formidable challenge :)

[paracycle]

I'd much rather see size explicitly handled separately, since it is not a calculation method. Moreover, the signature generated for size in this PR is not correct, since size does not take any parameters like column_name nor does it take a block.

[amomchilov]

Regarding this:

And when looking into how the .to_a worked, I'm now finding that the GeneratedRelationMethods#where chain typing is actually incorrect. If you give an argument like User.where(foo: 1) you do not in fact get to use the methods on the PrivateAssociationRelationWhereChain returned. Additionally, you can't call :to_a on an object that is a where chain (User.where.not.to_a), unlike what the generated rbi suggests.

Ufuk has opened a PR to make "where chains" no longer be relations: #2070

[amomchilov]

I'm OK with merging this as an incremental improvement over the current incorrect types. Could you please investigate the CI failures and see if they'r related to your changes?

[marknuzz]

I've been extremely busy lately so I may not have time to get to it this week but will do what I can