fix: handle indexes with a hidden compound

[BuonOmo]

This fixes regional related parts, hence fixes #344

[mmalek-sa]

This is not specific to CRDB >= 24.0.2. On 24.1.2 unique constraints on a region by row table will combine the crdb_region and column on the dumped schema. Since attishidden is not available in CRDB 24.1.2 this seems to do the trick:

def column_names_from_column_numbers(table_oid, column_numbers)
  Hash[query(<<~SQL, "SCHEMA")].values_at(*column_numbers).compact
    SELECT a.attnum, a.attname
    FROM pg_attribute a
    WHERE a.attrelid = #{table_oid}
    AND a.attnum IN (#{column_numbers.join(", ")})
    AND a.attisdropped <> 't'
    AND a.attname <> 'crdb_region'
  SQL
end

Without attisdropped <> t, I see weird ....pg.dropped.18... strings in the schema dump. All this is happening after upgrade to Rails 7.1.x and Cockroach AR Adapter 7.1.x. I can't reproduce it on Rails 7.0.x.

When it comes to unique_constraints, I also see a problem with indexes with expressions. For example:

    t.index "lower(name::STRING) ASC", name: "index_table_name_redacted_on_name_unique", unique: true

is becoming:

    t.unique_constraint [], name: "index_table_name_redacted_on_name_unique"

[mmalek-sa]

I created an issue with a workaround for unique_constraint here.

[BuonOmo]

@mmalek-sa I'm not targetting 24.0.2 but 24.2.2, which is the version where the related issue started raising. We do not need this patch before (hence the conditionnal!)

Also crdb_region is not the only issue, see the tests in this PR they are targetting hash sharded indexes!

---

EDIT: I understand now there is a second issue. I'll answer in the issue, and I'd rather solve it in a different PR so that we can ship this one quicker as it is blocking migrations for some users of 24.2

[rafiss]

nice, thanks for identifying the fix!