Optimize SELECT UNNEST in lateral joins

[taniabogatsch]

This PR adds the unnest_rewriter to the optimizer. While the two queries below are semantically equal, their query plans currently differ. The second query performs a duplicate eliminated join (DelimJoin), which is the default approach for lateral joins.

# UNNEST in the projection
SELECT i, UNNEST(i) AS j FROM (VALUES ([1, 2, 3]), ([4, 5])) t(i);
# different query plan, same result
SELECT * FROM (VALUES ([1, 2, 3]), ([4, 5])) t(i), (SELECT UNNEST(i)) t2(j);

However, a DelimJoin is not required for UNNESTs in the FROM clause. Instead, we can eliminate the DelimJoin. Then, we make the UNNEST the direct child of the former parent of the DelimJoin. Together with #5982, this PR tackles the performance issues in #5827.

Here are two query plans for comparison.

EXPLAIN SELECT * FROM (VALUES ([1, 2, 3])) t(i), (SELECT UNNEST(i)) t2(j);

Current master branch.

This PR.

[Mytherin]

Thanks for the PR! Looks good! Great performance gains.

I have been playing around with this for a bit and found this query does not work yet:

SELECT UNNEST(j) FROM (VALUES ([[1, 2, 3]]), ([[4, 5]])) t(i), (SELECT UNNEST(i)) t2(j);

Could you have a look at that?

Looks good to me otherwise.

[taniabogatsch]

Thanks for the find! I updated the implementation to skip unknown parent operators instead of throwing an internal exception, so this query works now! :)

[Mytherin]

Thanks for the changes!