[CALCITE-500] Fix EnumerableJoin to build the left side and probe the right

[atris]

No description provided.

[atris]

@julianhyde Can you please help review? Thanks

[zinking]

you will need a test, and some descriptions.

[michaelmior]

Also, the two commits should be squashed together.

[atris]

Ack, will fix this. I raised some questions on the dev list for the PR, will update post the resolution on that. Thanks for reviewing this PR, Michael and Zhen

…

On Tue, May 22, 2018 at 8:19 PM, Michael Mior ***@***.***> wrote: Also, the two commits should be squashed together. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#691 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABpO420O6iF8bDCIOavIGzoBMf1jwDxrks5t1CVlgaJpZM4UC2zr> .

-- Regards, Atri *l'apprenant*

[zabetak]

I haven't looked a lot in this issue but the fix seems wrong. I think it breaks LEFT, RIGHT, and FULL outer joins.

[atris]

Yeah, it isn't correct. I need to fix this sometime.

…

On Tue, Sep 25, 2018 at 11:20 PM Stamatis Zampetakis < ***@***.***> wrote: I haven't looked a lot in this issue but the fix seems wrong. I think it breaks LEFT, RIGHT, and FULL outer joins. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#691 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABpO46QmkpMCP_6l-ZIcREtmODt-pC6wks5uemz4gaJpZM4UC2zr> .

-- Regards, Atri *l'apprenant*

[vlsi]

Yes, it is not valid to just swap rightExpression and leftExpression.
leftResult.physType.generateAccessor(leftKeys) and friends should be replaced as well.

[atris]

Thanks for the pointer.

…

On Tue, Sep 25, 2018 at 11:23 PM Vladimir Sitnikov ***@***.***> wrote: Yes, it is not valid to just swap rightExpression and leftExpression. leftResult.physType.generateAccessor(leftKeys) and friends should be replaced as well. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#691 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABpO49YK8tGKOcKoG4cPa-fPtzrePqqCks5uem2IgaJpZM4UC2zr> .

-- Regards, Atri *l'apprenant*

[zabetak]

I would be inclined to change the cost function so that left and right inputs are set correctly. That way the implement method would not need to be changed at all.

[vlsi]

@zabetak , that depends on the background :)

For instance, in Oracle DB the first input is used to build the table.

[hsyuan]

@atris Are you going to continue working on this issue?

[vlsi]

Executor implementation has to be consistent with the planner.
For instance, the planner assumes the join would retain left input collation, so we can't use left input for lookup.

That is why I decline this PR in favour of adjusting the cost so it prefers smaller inputs on the right input.

See #1712