[
https://issues.apache.org/jira/browse/KYLIN-3249?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vsevolod Ostapenko updated KYLIN-3249:
--------------------------------------
Description:
Hybrid cubes are assigned default priority lower than regular cubes, which
leads to incorrect selection of a hybrid cube while a regular non-hybridized
cube with lower cost is available.
For example, model has a wide cube with full set of metrics and narrower cube
with top-N entries for a subset of metrics.
If wide cube is hybridized (due to a new metric addition), but top-N cube
remains unchanged and non-hybridized, top-N cube will be no longer queried,
causing query performance degradation.
The issue can be tracked to the
query/src/main/java/org/apache/kylin/query/routing/Candidate.java, where hybrid
cubes are assigned priority 0, while regular cubes are assigned priority of 1.
This unconditional priority assignment is incorrect as it only holds for cases
when there is only one cube "type" in the model or when all the cubes are
hybridized at the same time.
Simplest fix is to have hybrid priority to be the same as of a regular cube.
Plus, as an enhancement to the cube selection algorithm a new rule can be
implemented that will filter out regular candidate cubes that are included into
candidate hybrid cubes.
was:
Hybrid cubes are assigned default priority lower than regular cubes, which
leads to incorrect selection of a hybrid cube while a regular non-hybridized
cube with lower cost is available.
For example, model has a wide cube with full set of metrics and narrower cube
with top-N entries for a subset of metrics.
If wide cube is hybridized (due to a new metric addition), but top-N cube
remains unchanged and non-hybridized, top-N cube will be no longer queried,
'causing query performance degradation.
The issue can be tracked to the
query/src/main/java/org/apache/kylin/query/routing/Candidate.java, where hybrid
cubes are assigned priority 0, while regular cubes are assigned priority of 1.
This unconditional priority assignment is incorrect as it only holds for cases
when there is only one cube "type" in the model or when all the cubes are
hybridized at the same time.
Simplest fix is to have hybrid priority to be the same as of a regular cube.
Plus, as an enhancement to the cube selection algorithm a new rule can be
implemented that will filter out regular candidate cubes that are included into
candidate hybrid cubes.
> Default hybrid cube priority should be the same as of a regular cube
> --------------------------------------------------------------------
>
> Key: KYLIN-3249
> URL: https://issues.apache.org/jira/browse/KYLIN-3249
> Project: Kylin
> Issue Type: Bug
> Components: Query Engine
> Affects Versions: v2.2.0
> Environment: HDP 2.5.6, Kylin 2.2
> Reporter: Vsevolod Ostapenko
> Priority: Major
>
> Hybrid cubes are assigned default priority lower than regular cubes, which
> leads to incorrect selection of a hybrid cube while a regular non-hybridized
> cube with lower cost is available.
> For example, model has a wide cube with full set of metrics and narrower cube
> with top-N entries for a subset of metrics.
> If wide cube is hybridized (due to a new metric addition), but top-N cube
> remains unchanged and non-hybridized, top-N cube will be no longer queried,
> causing query performance degradation.
> The issue can be tracked to the
> query/src/main/java/org/apache/kylin/query/routing/Candidate.java, where
> hybrid cubes are assigned priority 0, while regular cubes are assigned
> priority of 1.
> This unconditional priority assignment is incorrect as it only holds for
> cases when there is only one cube "type" in the model or when all the cubes
> are hybridized at the same time.
> Simplest fix is to have hybrid priority to be the same as of a regular cube.
> Plus, as an enhancement to the cube selection algorithm a new rule can be
> implemented that will filter out regular candidate cubes that are included
> into candidate hybrid cubes.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
