[
https://issues.apache.org/jira/browse/HIVE-30182?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Denys Kuzmenko updated HIVE-30182:
----------------------------------
Description:
LlapDecider raises the minimum reducers of an auto-reduced vertex to
hive.tez.llap.min.reducer.per.executor (0.33) x
the cluster's executors (adjustAutoParallelism), so auto-reduce does not shrink
a reducer to a handful of tasks on a
large cluster. It applies the floor only to a ReduceWork: a reduce-side merge
join is a MergeJoinWork, whose main and
merged reduce works feed one vertex through their edges, and is skipped.
TPC-DS q51 at 10 TB on 24 LLAP daemons (288 executors): the FULL OUTER merge
join vertex ran 13 tasks while its input
reducers and the PTF after it ran 253; the floor would have kept it at 96 or
more.
Fix: apply the floor to the merge join's main work and each merged reduce work
(their edge properties are the ones
the vertex's edges use). TestLlapDecider covers it.
> Parquet: a leaf without its runtime value drops the whole
> ---------------------------------------------------------
>
> Key: HIVE-30182
> URL: https://issues.apache.org/jira/browse/HIVE-30182
> Project: Hive
> Issue Type: Sub-task
> Components: llap, Query Planning
> Reporter: Denys Kuzmenko
> Priority: Major
>
> LlapDecider raises the minimum reducers of an auto-reduced vertex to
> hive.tez.llap.min.reducer.per.executor (0.33) x
> the cluster's executors (adjustAutoParallelism), so auto-reduce does not
> shrink a reducer to a handful of tasks on a
> large cluster. It applies the floor only to a ReduceWork: a reduce-side merge
> join is a MergeJoinWork, whose main and
> merged reduce works feed one vertex through their edges, and is skipped.
> TPC-DS q51 at 10 TB on 24 LLAP daemons (288 executors): the FULL OUTER merge
> join vertex ran 13 tasks while its input
> reducers and the PTF after it ran 253; the floor would have kept it at 96 or
> more.
> Fix: apply the floor to the merge join's main work and each merged reduce
> work (their edge properties are the ones
> the vertex's edges use). TestLlapDecider covers it.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
