[
https://issues.apache.org/jira/browse/KYLIN-6042?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17928343#comment-17928343
]
Guoliang Sun commented on KYLIN-6042:
-------------------------------------
h3. RootCause
When recommending model indexes, two modelContexts are recommended with only
two differences:
1. Whether the indexRexItemMap has a value, which corresponds to the detailed
index column number in the issue.
2. The snapShotSelected attribute value, where true indicates the information
is from a snapshot, and false indicates it is from the model.
Looking at the SQL, normally it could be handled within one modelContext.
However, due to special function processing in the join condition columns and
the use of a right join, it is split into two modelContexts. One attempts to
match the snapshot, while the other matches the model. When a snapshot is
identified, the detailed index is not included in the recommendation; the index
is recommended only on the model.
{code:java}
SELECT
c.*
FROM
(
select
a.*,
b.*
from
(
SELECT
LO_ORDERPRIOTITY AS LO_ORDERPRIOTITY
FROM
SSB.LINEORDER
) a
right join (
select
*
from
(
SELECT
trim(C_NAME) AS C_NAME
FROM
SSB.CUSTOMER
) d
) b on trim(a.LO_ORDERPRIOTITY) = b.C_NAME {code}
In subsequent processing logic, only the index information in the first found
modelContext (which is the snapshot) is processed. The modelId is used when
searching for the model, but this value is the same in both modelContexts.
Following the operation order described in the reproduction steps, the first
found modelContext is consistently the one without index information (i.e., the
snapshot). As a result, the deletion operation becomes ineffective, leading to
the observed behavior where the parameters appear to have no effect.
> Model acceleration interface: The parameter discard_table_index is not taking
> effect
> ------------------------------------------------------------------------------------
>
> Key: KYLIN-6042
> URL: https://issues.apache.org/jira/browse/KYLIN-6042
> Project: Kylin
> Issue Type: Bug
> Affects Versions: 5.0.0
> Reporter: Guoliang Sun
> Priority: Major
>
> 1. When the parameter discard_table_index is true, detailed index suggestions
> are still generated. API: kylin/api/models/model_optimization
> 2. SQL reference
> {code:java}
> SELECT c.* FROM (
> select a.*, b.* from (
> SELECT LO_ORDERPRIOTITY AS LO_ORDERPRIOTITY
> FROM HANHAN_LOGICAL_VIEW.LOGICAL_VIEW_LINEORDER
> ) a
> right join (
> select * from (
> SELECT trim(C_NAME) AS C_NAME
> FROM SSB.CUSTOMER
> ) d
> ) b
> on trim(a.LO_ORDERPRIOTITY)=b.C_NAME
> ) c{code}
> 3. Two modelContexts were generated with the same targetModel. The
> findFirst() here caused the IndexRexItemMap in the second modelContext not to
> be removed, resulting in the successful creation of detailed index
> suggestions.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
