[
https://issues.apache.org/jira/browse/HIVE-23965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17176161#comment-17176161
]
Stamatis Zampetakis commented on HIVE-23965:
--------------------------------------------
Before finalizing this case it may be good to decide if we want to keep or
delete the old TPC-DS drivers (or better say tests) namely
TestTezPerfCliDriver, TestTezPerfConstraintsCliDriver as part of this issue.
A few advantages/disadvantages of keeping the old tests are outlined below.
+Advantages:+
* Better code coverage. Due to differences in stats, and configurations we end
up with different plans so potentially we are covering different codepaths.
* The old drivers can be run on local dev environment without the need of
installing Docker.
+Disadvantages+
* Maintenance cost. Now for every change in the planner we may need to update
the results from three drivers (~300 queries).
* Unrealistic plans. As mentioned previously the table stats are not obtained
from a single TPC-DS scale factor so we in some cases we may never see this
plans in practice; this can be seen also as a bug that we could possibly fix.
* Test execution time. Obviously running three test suites instead of one is
going to take more time and there is nothing that we can do about it.
In case there is disagreement we can keep them for now and postpone the
decision in another JIRA. Let me know what you think.
> Improve plan regression tests using TPCDS30TB metastore dump and custom
> configs
> -------------------------------------------------------------------------------
>
> Key: HIVE-23965
> URL: https://issues.apache.org/jira/browse/HIVE-23965
> Project: Hive
> Issue Type: Improvement
> Reporter: Stamatis Zampetakis
> Assignee: Stamatis Zampetakis
> Priority: Major
> Labels: pull-request-available
> Time Spent: 1h
> Remaining Estimate: 0h
>
> The existing regression tests (HIVE-12586) based on TPC-DS have certain
> shortcomings:
> The table statistics do not reflect cardinalities from a specific TPC-DS
> scale factor (SF). Some tables are from a 30TB dataset, others from 200GB
> dataset, and others from a 3GB dataset. This mix leads to plans that may
> never appear when using an actual TPC-DS dataset.
> The existing statistics do not contain information about partitions something
> that can have a big impact on the resulting plans.
> The existing regression tests rely on more or less on the default
> configuration (hive-site.xml). In real-life scenarios though some of the
> configurations differ and may impact the choices of the optimizer.
> This issue aims to address the above shortcomings by using a curated
> TPCDS30TB metastore dump along with some custom hive configurations.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
