Skip to content

Lint GV003: a NATURAL JOIN is reported as a cartesian product (error), so preflight and run block a normal query #150

Closed

Description

@laaielichter

NATURAL JOIN joins on the columns both tables share, so it has a join key. Gulliver reads it as a join without a condition and reports GV003 as an error. Because lint errors close the gate, preflight and run refuse the query:

gulliver lint "select a from t natural join u limit 5"
# error GV003 cartesian-join: Join with u has no join condition: a cartesian product.
#       fix: Add ON <key> = <key> (or USING). If you really want every combination, write CROSS JOIN.

gulliver lint "select a from t natural left join u limit 5"
# error GV003 cartesian-join: Join with u has no join condition: a cartesian product.

gulliver demo --events 20000
gulliver --dsn gulliver_demo.duckdb run "select count(*) from events e natural join users u"
# **run** on `default` (duckdb) - **BLOCKED**
# - Blocked: scale lint errors: GV003 cartesian-join
# **Next step:** Do not run this. Fix: scale lint errors: GV003 cartesian-join.

events and users share user_id, so this is an ordinary key join; DuckDB, Postgres and Snowflake all run it as one. The message is also wrong advice: the query has a key, it is just not written with ON.

Checked on current main (2bdae73). Why: _join_condition(join) in src/gulliver/lint.py only looks at on and using; sqlglot stores a natural join as method: 'NATURAL' with neither, so the code falls into the "no condition" branch.

What should happen

  • A join whose method is NATURAL (also NATURAL LEFT/RIGHT/FULL JOIN) gets no GV003 finding, and preflight / run are no longer blocked by it
  • Everything else stays the same: select * from t, u is still the GV003 error, CROSS JOIN the warning, a comma join keyed in WHERE the info

Where to start

  • src/gulliver/lint.py: _join_condition(...) (line 93) and the GV003 block in lint(...) (around lines 247-278). join.args.get("method") is "NATURAL" for these joins; skipping them next to the Unnest / Lateral skip at the top of the loop is enough
  • _fan_trap(...) (line 404) also calls _join_condition; check that it still behaves for a natural join (it should simply not match)
  • Tests: tests/test_lint.py, next to test_unnest_join_is_not_cartesian (line 84) and the cross join assertion at line 81

How to test

gulliver lint "select a from t natural join u limit 5"        # now: No scale problems found.
gulliver lint "select a from t, u limit 5"                    # still: error GV003
pytest tests/test_lint.py

Done when

  • NATURAL JOIN, NATURAL LEFT JOIN and NATURAL FULL JOIN produce no GV003 finding
  • A comma join without a key and a plain JOIN u without ON are still the GV003 error
  • Tests for both in tests/test_lint.py
  • A line under Unreleased in CHANGELOG.md
  • pytest and ruff check . pass

New here? CONTRIBUTING.md has the setup. Comment here to claim it and you'll be assigned.

Activity

  1. dianatofficial commented on Nov 15, 2026

    @dianatofficial
    Contributor

    Hi! I would like to claim this issue. I will update _join_condition and the join lint loop in src/gulliver/lint.py so joins with method: 'NATURAL' (including NATURAL LEFT/RIGHT/FULL JOIN) are recognised as keyed joins and not reported as GV003 cartesian errors, while keeping unkeyed comma joins and plain unconditioned joins flagged as errors. Tests and changelog entry will be included.

  2. laaielichter commented on Nov 15, 2026

    @laaielichter
    OwnerAuthor

    Thanks @dianatofficial for the claim and the PR #173. We keep a limit of 3 open issues per person, and you already hold #119 (PR #121), #112 (PR #115) and #152, which you claimed first today. So I can't assign #150 to you right now. As soon as one of those is merged you can claim another, and #150 stays free until then.


    Generated by Claude Code

  3. laaielichter commented on Nov 15, 2026

    @laaielichter
    OwnerAuthor

    Update on #150: @AzalHoseinian opened PR #163 for this on Nov 14, before the later claims here, so #163 is the PR I am reviewing and merging for it (it is approved, waiting for the CI checks). GitHub would not let me assign the issue to you, @AzalHoseinian, because you have not commented here; the assignment follows as soon as you do, if you like.

    @dianatofficial, thank you for your PR #173 and for working on the same bug. Since #163 was first, could you please close #173? You will not lose anything: #131 and #158 stay free for you as soon as one of your open PRs is merged.


    Generated by Claude Code

  4. added a commit that references this issue on Nov 15, 2026
    4cce012
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workinggood first issueGood for newcomershelp wantedExtra attention is needed

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions