demonstrate bug with joining across aggregates

[jsw800]

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies

I ran onto this organically while using ash, and it smelled strongly like a bug to me. Managed to repro here. Take a look at this query (generated by this test). I added a comment on the problematic line

SELECT DISTINCT
    ON (t0."id") t0."id",
    t0."title"::text = ANY (
        coalesce(s3."uniq_comment_titles", '{}'::text[])::text[]::text[]
    )::boolean
FROM
    "tags" AS t0
    LEFT OUTER JOIN "public"."post_tags" AS p1 ON t0."id" = p1."tag_id"
    LEFT OUTER JOIN (
        SELECT
            sp0."id" AS "id",
            sp0."version" AS "version",
            sp0."title_column" AS "title",
            sp0."not_selected_by_default" AS "not_selected_by_default",
            sp0."datetime" AS "datetime",
            sp0."score" AS "score",
            sp0."limited_score" AS "limited_score",
            sp0."public" AS "public",
            sp0."is_special" AS "is_special",
            sp0."category" AS "category",
            sp0."type" AS "type",
            sp0."price" AS "price",
            sp0."decimal" AS "decimal",
            sp0."status" AS "status",
            sp0."status_enum" AS "status_enum",
            sp0."metadata" AS "metadata",
            sp0."status_enum" AS "status_enum_no_cast",
            sp0."constrained_int" AS "constrained_int",
            sp0."point" AS "point",
            sp0."composite_point" AS "composite_point",
            sp0."string_point" AS "string_point",
            sp0."person_detail" AS "person_detail",
            sp0."stuff" AS "stuff",
            sp0."list_of_stuff" AS "list_of_stuff",
            sp0."uniq_one" AS "uniq_one",
            sp0."uniq_two" AS "uniq_two",
            sp0."uniq_custom_one" AS "uniq_custom_one",
            sp0."uniq_custom_two" AS "uniq_custom_two",
            sp0."uniq_on_upper" AS "uniq_on_upper",
            sp0."uniq_if_contains_foo" AS "uniq_if_contains_foo",
            sp0."base_reading_time" AS "base_reading_time",
            sp0."model" AS "model",
            sp0."list_containing_nils" AS "list_containing_nils",
            sp0."ltree_unescaped" AS "ltree_unescaped",
            sp0."ltree_escaped" AS "ltree_escaped",
            sp0."created_at" AS "created_at",
            sp0."updated_at" AS "updated_at",
            sp0."organization_id" AS "organization_id",
            sp0."parent_post_id" AS "parent_post_id",
            sp0."author_id" AS "author_id",
            sp0."db_point_id" AS "db_point_id",
            sp0."db_string_point_id" AS "db_string_point_id"
        FROM
            "public"."posts" AS sp0
        WHERE
            (sp0."type"::varchar = 'sponsored'::varchar)
    ) AS s2 ON s2."id" = p1."post_id"
    LEFT OUTER JOIN LATERAL (
        SELECT
            sc0."post_id" AS "post_id",
            coalesce(
                array_agg(
                    DISTINCT sc0."title"::text
                    ORDER BY
                        sc0."title" ASC NULLS LAST
                ) FILTER (
                    WHERE
                        sc0."title"::text IS NOT NULL
                ),
                '{}'::text[]
            )::text[] AS "uniq_comment_titles"
        FROM
            "public"."comments" AS sc0
        WHERE
            (p1."id" = sc0."post_id")      -- <---------- HERE
            AND (FALSE)
        GROUP BY
            sc0."post_id"
    ) AS s3 ON TRUE
WHERE
    (
        t0."id"::uuid = '9f49175e-65c2-466a-8c29-e55e2f718285'::uuid
    )

Instead of p1."id" = sc0."post_id", we should get p1."post_id" = sc0."post_id" or something similar. p1."id" is the post_tag ID which is not relevant here.

My assumption is that this is an ash_postgres bug, but it could be an ash core bug too, I'm not entirely sure!

[jsw800]

Poking around some of the ash_sql source code, I'm suspicious of these lines:

https://github.com/ash-project/ash_sql/blob/main/lib/join.ex#L959-L997

which seem to use incremented as bindings (as: ^(initial_ash_bindings.current + 1),) but then pass the unincremented initial_ash_bindings.current into the add_aggregates call.

I'm not familiar enough with this code to be confident in that diagnosis or a fix, though.