Laurent Goujon created CALCITE-2538:
---------------------------------------
Summary: Missing RelMetadataQuery invalidation before subset cost
improvement propagation
Key: CALCITE-2538
URL: https://issues.apache.org/jira/browse/CALCITE-2538
Project: Calcite
Issue Type: Bug
Components: core
Reporter: Laurent Goujon
Assignee: Julian Hyde
When planner log level is set to debug, several tests are failing with the
following error message:
{noformat}
rel
[rel#11408:EnumerableProject.ENUMERABLE.[](input=rel#11180:Subset#2.ENUMERABLE.[0],DUMMY=0)]
has lower cost {24.5 rows, 39.5 cpu, 0.0 io} than best cost {28.0 rows, 36.0
cpu, 0.0 io} of subset [rel#11177:Subset#3.ENUMERABLE.[]]
{noformat}
The following query (from 'sql/agg.iq') should exhibit the issue:
{code:sql}
select sum(e.sal) as s from "scott".emp e join "scott".emp m on e.mgr = e.empno;
{code}
I tried to trace it back and my understanding is that during cost improvement
propagation, a subset has its cost/best node changed, and it triggers
recomputing the cost of the parent subset, but the current best node of the
parent has its value cached in the current rel metadata query, and is not
recomputed. During a later validation, this error is detected and the assertion
is thrown.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
