[jira] [Commented] (IMPALA-7769) Handle CAST(NULL AS type) in rewrites

["Paul Rogers (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/IMPALA-7769?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16679202#comment-16679202 ] 

Paul Rogers commented on IMPALA-7769:
-------------------------------------

An example failure in {{PlannerTest.testEmpt()}}, for the following case:

{noformat}
# IMPALA-1860: INSERT/CTAS should evaluate and apply constant predicates.
with t as (select * from functional.alltypes where coalesce(NULL) > 10)
insert into functional.alltypes partition(year, month)
select * from t
{noformat}

Monitor the rewrite output. We get:

{noformat}
Before: coalesce(NULL) > 10
After:  NULL > 10
{noformat}

Here, the {{NULL}} on the second line is actually {{CAST(NULL AS TINYINT)}}. The expression should have been further rewritten to just {{NULL}}. The problem is that the constant folding rule looks for expressions that contains literals, but here is needs to also look for a CAST of a NULL literal.

> Handle CAST(NULL AS type) in rewrites
> -------------------------------------
>
>                 Key: IMPALA-7769
>                 URL: https://issues.apache.org/jira/browse/IMPALA-7769
>             Project: IMPALA
>          Issue Type: Improvement
>          Components: Frontend
>    Affects Versions: Impala 3.0
>            Reporter: Paul Rogers
>            Priority: Minor
>
> Consider the following query:
> {code:sql}
> SELECT IFNULL(NULL + 1, id) FROM alltypessmall;
> {code}
> Visualize the rewritten query after analysis (using the new {{FullRewriteTest}}.) We get:
> {code:sql}
> SELECT CASE WHEN CAST(NULL AS INT) IS NULL THEN id ELSE NULL END FROM alltypessmall
> {code}
> (This is what the expression actually contains. The {{toSql()}} method lies and says that the statement is:
> {code:sql}
> SELECT CASE WHEN NULL IS NULL THEN id ELSE NULL END FROM alltypessmall
> {code}
> which causes confusion when debugging.)
> Expected:
> {code:sql}
> SELECT id FROM alltypessmall
> {code}
> Another case:
> {code:sql}
> CASE WHEN NULL + 1 THEN 10 ELSE 20 END
> {code}
> Ensure the {{NULL + 1}} is constant folded with a cast. Then, in {{CASE}}:
> {code:java}
>     for (int i = loopStart; i < numChildren - 1; i += 2) {
>       if (expr.getChild(i).isLiteral()) {
>         canSimplify = true;
>         break;
>       }
>     }
> {code}
> The {{isLiteral()}} method won’t match the cast and the simplification won’t fire.
> The reason is that there are multiple rewrites, one of which has a flaw.
> # Rewrite {{IFNULL to CASE}}
> # Rewrite {{NULL + 1}} to {{CAST(NULL AS SMALLINT)}}
> # Try to rewrite {{CAST(NULL AS SMALLINT) IS NULL}}, but fail because CAST is not a literal.
> In addition to the {{CASE}} issue above, the {{FoldConstantsRule}} itself is tripped up. It is supposed to simplify the {{... IS NULL}} expression above, but does not.
> The code in question in {{FoldConstantsRule.apply()}} is:
> {code:java}
>     for (Expr child: expr.getChildren()) if (!child.isLiteral()) return expr;
> {code}
> In fact, this check is too restrictive. Need a new {{isLiteralLike()}} which should work like {{IsNullLiteral()}}:
> * True if the node is a literal.
> * True if the node is a cast of a literal.
> (Can't change {{isLiteral()}} since there are places that assume that this existing predicate indicates that the node is exactly a {{LiteralExpr}}.)
> Impala already has a predicate that does what is needed: {{isConstant()}}. However, the code in {{FoldConstantsRule.apply()}} specifically excludes calling it:
> {code:java}
>     // Avoid calling Expr.isConstant() because that would lead to repeated traversals
>     // of the Expr tree. Assumes the bottom-up application of this rule. Constant
>     // children should have been folded at this point.
> {code}
> The new method solves the repeated traversal problem. With it, the test query now simplifies to the expected result.



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

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscribe@impala.apache.org
For additional commands, e-mail: issues-all-help@impala.apache.org