Regression in `map/2` with left join returning `%{field: nil}` instead of `nil` in Ecto 3.13.1

[halfdan]

Elixir version

Elixir 1.18.2 (compiled with Erlang/OTP 27)

Database and Version

PostgreSQL 16

Ecto Versions

3.13.1

Database Adapter and Versions (postgrex, myxql, etc)

postgrex

Current behavior

In 3.13.x the test (see expected behaviour) returns a struct with all keys present, but values set to nil.

     test/asset_test.exs:35
     Assertion with == failed
     code:  assert result == {asset.id, nil}
     left:  {asset.id, %{id: nil, name: nil}}
     right: {asset.id, nil}

Expected behavior

Unclear - but it's a breaking change.

The following test passes in 3.12.x:

test "Ecto 3.12.x: left join map/2 returns nil when join is nil" do
  asset = insert(:file)

  result =
    Asset
    |> from(as: :asset)
    |> join(:left, [asset: a], parent in Asset, on: a.parent_id == parent.id, as: :parent)
    |> where([asset: a], is_nil(a.parent_id) and a.id == ^asset.id)
    |> select([a, p], {a.id, map(p, [:id, :name])})
    |> Repo.one!()

  assert result == {asset.id, nil}
end

(the left joined parent does not exist so is nil)

[josevalim]

Yes, this is an intentional change we did on Ecto v3.13 to solve a consistency bug. Although I think I did not add this to the CHANGELOG, right @greg-rychlewski? It mentions the parent select_merge issue but not the changes to map?

[greg-rychlewski]

@josevalim Yeah I remember this one started out with select_merge but then we discovered it actually extended to any kind of join with map/structs. And then we decided to make the maps consistent and leave structs as is. This comment and the one after by you were the conclusions.