You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@calcite.apache.org by "Haisheng Yuan (JIRA)" <ji...@apache.org> on 2019/05/15 03:38:00 UTC

[jira] [Commented] (CALCITE-2624) Add a rule to copy a sort below a join operator

    [ https://issues.apache.org/jira/browse/CALCITE-2624?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16839978#comment-16839978 ] 

Haisheng Yuan commented on CALCITE-2624:
----------------------------------------

The PR is mixing physical property enforcement with logical plan exploration. The physical property should only be requested by physical operator by enforcing expected sort order. The patch is creating more useless plan alternatives than necessary. It only makes sense to push a top limit into the outer relation of an outer join, other than that, I don't think it is the right way to go.

> Add a rule to copy a sort below a join operator
> -----------------------------------------------
>
>                 Key: CALCITE-2624
>                 URL: https://issues.apache.org/jira/browse/CALCITE-2624
>             Project: Calcite
>          Issue Type: Improvement
>          Components: core
>    Affects Versions: 1.17.0
>            Reporter: Stamatis Zampetakis
>            Assignee: Khawla Mouhoubi
>            Priority: Major
>              Labels: pull-request-available
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> Currently, the only rule that allows a sort to traverse a binary operator is the SortJoinTransposeRule. The rule was introduced mainly to push limits in the case of left and right outer joins (see CALCITE-831).
> I assume that the main reason that we don't have more rules is that sorts with limits and offsets cannot be pushed safely below many types of join operators. However, in many cases, it is possible and beneficial for optimization purposes to just push the sort without the limit and offset. Since we do not know in advance if the join operator preserves the order we cannot remove (that is why I am saying copy and not transpose) the sort operator on top of the join. The latter is not really a problem since the SortRemoveRule can detect such cases and remove the sort if it is redundant.
> A few concrete examples where this optimization makes sense are outlined below:
>  * allow the sort to be later absorbed by an index scan and disappear from the plan (Sort + Tablescan => IndexScan with RelCollation);
>  * allow operators that require sorted inputs to be exploited more easily (e.g., merge join);
>  * allow the sort to be performed on a possibly smaller result (assuming that the physical binary operator that is going to be used preserves the order of left/right input and the top sort operator can be removed entirely).
> I propose to add a new rule (e.g., SortCopyBelowJoinRule, SortJoinCopyBelowRule) which allows a sort to be copied to the left or right (or to both if it is rather easy to decompose the sort) of a join operator (excluding the limit and offset attributes) if the respective inputs are not already sorted. 



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)