[
https://issues.apache.org/jira/browse/IMPALA-15282?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksandr Efimov updated IMPALA-15282:
--------------------------------------
Component/s: Frontend
> Calcite planner: test Substrait export of the pre-conversion logical plan
> -------------------------------------------------------------------------
>
> Key: IMPALA-15282
> URL: https://issues.apache.org/jira/browse/IMPALA-15282
> Project: IMPALA
> Issue Type: Improvement
> Components: Frontend
> Reporter: Aleksandr Efimov
> Assignee: Aleksandr Efimov
> Priority: Major
>
> IMPALA-15237 added a direct Calcite-only harness that exposes the optimized
> logical RelNode immediately before Impala conversion. That plan is standard
> Calcite, so it can be handed to other tools; the check that keeps it that way
> is whether it still converts cleanly to Substrait.
>
> Add a test-only compatibility test that takes the same pre-conversion plan,
> converts it with Isthmus, serializes it to protobuf, parses it back, and
> asserts on what a consumer would read: output schema, field references, and
> the relation and function inventories. The client-visible labels are Impala's
> own, so the test puts them on the Substrait root and checks that they survive
> the round trip.
>
> Covered shapes: a virtual table (VALUES) under a Project, a named read with
> Filter and Project, scalar arithmetic, and a relation Isthmus cannot convert,
> which has to fail with a typed error instead of publishing a partial plan.
>
> The Isthmus dependency is test-scoped: no production dependency, no exporter,
> no query option, nothing added to impala-package. The test fails when a change
> at the seam makes one of the covered plans unconvertible or changes what a
> consumer would read - a custom operator in place of a standard one, a custom
> RelNode, a different type or a different field reference.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
