[
https://issues.apache.org/jira/browse/TEZ-2956?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15018391#comment-15018391
]
Bikas Saha commented on TEZ-2956:
---------------------------------
This is another case which should be covered via TEZ-2943 where we overhaul the
heuristics. However, this patch adds a test case which should continue to pass
after TEZ-2943. So lets take this patch.
On the patch - I dont think we should overrule minPartitions since this is a
user defined config. This may break the assumptions in the overall DAG plan
that users have submitted. Thoughts?
{code}
if(desiredTaskParallelism < minTaskParallelism) {
- desiredTaskParallelism = minTaskParallelism;
+ desiredTaskParallelism =
+ (totalNumBipartiteSourceTasks == 0) ? 1: minTaskParallelism;
}{code}
> Handle auto-reduce parallelism when the totalNumBipartiteSourceTasks is 0
> -------------------------------------------------------------------------
>
> Key: TEZ-2956
> URL: https://issues.apache.org/jira/browse/TEZ-2956
> Project: Apache Tez
> Issue Type: Bug
> Reporter: Rajesh Balamohan
> Assignee: Rajesh Balamohan
> Attachments: TEZ-2956.1.patch, TEZ-2956_DAG.png, With_Patch.png,
> Without_Patch.png
>
>
> In certain cases (e.g M --> R --> R), if the parent vertex has 0 tasks tez
> currently does not modify the parallelism factor in downstream.
> e.g
> {noformat}
> SELECT ss_store_sk,
> ss_sold_date_sk,
> ss_quantity,
> ss_sales_price,
> LEAD(ss_sales_price, 1) OVER(PARTITION BY ss_store_sk
> ORDER BY ss_quantity)
> FROM store_sales
> WHERE ss_sold_date_sk IS NOT NULL
> AND ss_quantity IS NOT NULL
> AND ss_sales_price > 2857684
> AND ss_sales_price < 2857685
> AND ss_store_sk > 10234233423
> AND ss_store_sk < 20234234324
> ORDER BY ss_store_sk,
> ss_sold_date_sk;
> {noformat}
> This would launch DAG "M1 (0) --> R2 (156) --> R3 (1)". However, R2 retains
> the parallelism of 156 even though no output would be generated in M1.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
