feat: Support swap for RightMark Join
#17651
Conversation
github-actions
bot
added
logical-expr
|
@comphead Decided to circle back to this after forgetting that this was reverted 😆 Are you able to run extended tests? |
|
I think we got something broken, you can still run extended tests on your fork, against your branch |
|
@comphead It is passing on my end. I added the slt test that is the query that caused the error from the previous pull request. |
|
cc @comphead I believe this should be good for another look! |
| .partition_statistics(None) | ||
| .unwrap() | ||
| .total_byte_size, | ||
| Precision::Inexact(2097152) |
There was a problem hiding this comment.
can this number be a reason of flaky test in the future?
There was a problem hiding this comment.
I don't think so, i have copied this from the other previous join selection swapping testing and thos ehave been working fine.
comphead
left a comment
comphead
left a comment
There was a problem hiding this comment.
Thanks @jonathanc-n I think the PR is good to go
Co-authored-by: Oleks V <comphead@users.noreply.github.com>
3 participants
Which issue does this PR close?
Part of fix for #16488
Changes
During decorrelation of an exists subquery into a LeftMark join, the inner side still carries the literal from 'select 1'. The schema looks like this:
right = Projection: [Int64(1), x.b]Projection Pushdown later drops the unused constant and rewrites the right child to just [x.b]. That “drop the first column” step shifts right-side column indices (from b@1 to b@0). During projection mapping it still assumes the old index (often doing “offset by left_width” and not accounting for the mark column), so when it re-indexes into the child it points at the wrong slot (which is what hit the error)
to fix this, I removed the projections when optimizer builds the join
Tests
The test is added to slt.