elixir-sqlite / ecto_sqlite3

An Ecto SQLite3 adapter.
https://hexdocs.pm/ecto_sqlite3
MIT License
300 stars 45 forks source link

Update query transformation issue #150

Open LostKobrakai opened 3 months ago

LostKobrakai commented 3 months ago

Elixir version

1.17.2

Database and Version

sqlite 3 (as brought by adapter)

Ecto Versions

3.12.1

Database Adapter and Versions (postgrex, myxql, etc)

exqlite 0.23, ecto_sqlite3 0.17.0

Current behavior

The below query doing left joins is turned into multiple FROM clauses, which according to the sqlite docs stating they modeled it after postgres and postgres documenting it to be a cross join, would be a different kind of join.

#Ecto.Query<from p0 in Planner.ProjectsQwer.Project,
 left_join: q1 in subquery(from q0 in "qwer_placeables",
  where: q0.project_id == ^1,
  group_by: [fragment("TRUE")],
  select: %{
  json:
    fragment(
      "json_group_array(?)",
      fragment(
        "json_object(?,?,?,?,?,?,?,?,?,?,?,?,?,?)", […fields]
      )
    )
}),
 as: :placeables, on: true,
 left_join: q2 in subquery(from q0 in "qwer_links",
  where: q0.project_id == ^1,
  group_by: [fragment("TRUE")],
  select: %{
  json:
    fragment(
      "json_group_array(?)",
      fragment(
        "json_object(?,?,?,?,?,?,?,?,?,?,?,?)", […fields]
      )
    )
}),
 as: :links, on: true, where: p0.id == ^1,
 update: [
  set: [
    snapshots:
      fragment(
        "json_insert(?,?,?)",
        fragment(
          "CASE WHEN ? THEN ? ELSE ? END",
          fragment("json_array_length(?)", p0.snapshots) >= 10,
          fragment("json_remove(?,?)", p0.snapshots, "$[0]"),
          p0.snapshots
        ),
        "$[#]",
        fragment("json_object(?,?,?,?)", "placeables", q1.json, "links", q2.json)
      )
  ]
]>
UPDATE "qwer_projects" AS q0 
SET "snapshots" = json_insert(CASE WHEN json_array_length(q0."snapshots") >= 10 THEN json_remove(q0."snapshots",'$[0]') ELSE q0."snapshots" END,'$[#]',json_object('placeables',s1."json",'links',s2."json")) 
FROM 
    (
        SELECT json_group_array(json_object([...fields])) AS "json" 
        FROM "qwer_placeables" AS sq0 
        WHERE (sq0."project_id" = ?) 
        GROUP BY TRUE
    ) AS s1, 
    (
        SELECT json_group_array(json_object([...fields])) AS "json" 
        FROM "qwer_links" AS sq0 
        WHERE (sq0."project_id" = ?) 
        GROUP BY TRUE
    ) AS s2 
    WHERE (q0."id" = ?)

Expected behavior

The join semantics are retained.

Follow up to https://github.com/elixir-ecto/ecto/issues/4502

warmwaffles commented 3 months ago

I think I kind of understand what you are asking.

Do you have a stub schema, some example rows, and the expected results update results?

In the mean time, I'll see if I can come up with a simpler reproduction of this problem without the json stuff. (although having that in the regression test wouldn't be a bad thing either)

LostKobrakai commented 3 months ago
Repo.insert_all("table", [%{number: 0}])
a = from x in fragment("json_each(?)", "[]"), select: %{value: x.value}

u = 
  from x in "table",
    left_join: a in subquery(a), on: true,
    update: [set: [number: coalesce(a.value, 1)]]

{1, _} = Repo.update_all(u, [])

This should be a quote simple example. Just needs a table with a number column. You could replace the subquery with a table having a value column (int), but being empty. Currently the update would result in no row updated, where it should.