[
https://issues.apache.org/jira/browse/IMPALA-15300?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksandr Efimov reassigned IMPALA-15300:
-----------------------------------------
Assignee: Aleksandr Efimov
> Calcite planner: narrow integer operands for Substrait substring and round
> --------------------------------------------------------------------------
>
> Key: IMPALA-15300
> URL: https://issues.apache.org/jira/browse/IMPALA-15300
> Project: IMPALA
> Issue Type: Sub-task
> Components: Frontend
> Reporter: Aleksandr Efimov
> Assignee: Aleksandr Efimov
> Priority: Major
>
> IMPALA-15297 maps Impala operators onto the Substrait functions they stand
> for by name. Two of them need more than a name.
> Substrait's substring takes its position and length as i32, and Impala
> declares substr with bigint parameters, so the call arrives as
> SUBSTR(varchar, i64, i64) and matches nothing. Round is the same mismatch the
> other way: Substrait's round takes the digit count as i32 while Impala types
> it from the literal, so a two-argument round arrives with an i8. Calcite's
> own SUBSTRING runs into this here as well, so the width is Substrait's choice
> rather than an Impala quirk, and Isthmus's own coercion takes the operand
> types as they come.
> Narrow those operands to i32 before matching. Isthmus has a hook for
> supplying different operands but it is package private; the converter itself
> is public, so the narrowing belongs there, at the point where the two type
> systems meet, rather than in a pass over the plan.
> The narrowing is an explicit cast rather than a silent reinterpretation,
> which Isthmus writes out with THROW_EXCEPTION behaviour. That is stricter
> than Impala, which returns no rows rather than failing for a position past
> the i32 range. Substrait has no wider substring to target, so the choice is
> between saying so in the plan and not exporting these queries at all.
> Over the TPC-DS queries that carry Calcite goldens this takes the exporter
> from 84 of 110 to 96: substr accounts for ten of them and round for two.
> A one-argument round stays out. Every Substrait round variant takes a digit
> count, so there is nothing to match, and carrying it across would mean
> inventing an operand rather than adjusting one.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
