[
https://issues.apache.org/jira/browse/HIVE-24485?focusedWorklogId=548959&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-548959
]
ASF GitHub Bot logged work on HIVE-24485:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 06/Feb/21 03:45
Start Date: 06/Feb/21 03:45
Worklog Time Spent: 10m
Work Description: okumin commented on a change in pull request #1744:
URL: https://github.com/apache/hive/pull/1744#discussion_r571351354
##########
File path: ql/src/java/org/apache/hadoop/hive/ql/parse/GenTezUtils.java
##########
@@ -81,6 +82,12 @@ public static ReduceWork createReduceWork(
boolean isAutoReduceParallelism =
context.conf.getBoolVar(HiveConf.ConfVars.TEZ_AUTO_REDUCER_PARALLELISM);
+ float slowStartMaxSrcFraction = context.conf.getFloat(
+ ShuffleVertexManager.TEZ_SHUFFLE_VERTEX_MANAGER_MAX_SRC_FRACTION,
+
ShuffleVertexManager.TEZ_SHUFFLE_VERTEX_MANAGER_MAX_SRC_FRACTION_DEFAULT);
+ float slowStartMinSrcFraction = context.conf.getFloat(
+ ShuffleVertexManager.TEZ_SHUFFLE_VERTEX_MANAGER_MIN_SRC_FRACTION,
+
ShuffleVertexManager.TEZ_SHUFFLE_VERTEX_MANAGER_MIN_SRC_FRACTION_DEFAULT);
Review comment:
I wonder if we should add new parameters and use them instead of ones
defined in ShuffleVertexManager.
Looking through this class, it would be consistent to introduce new
parameters and use `TEZ_SHUFFLE_VERTEX_MANAGER_{MIN, MAX}_SRC_FRACTION` only in
`DagUtils.java`.
However, Hive on Tez can access `TEZ_SHUFFLE_VERTEX_MANAGER_{MIN,
MAX}_SRC_FRACTION` in other cases. So I thought it might be also confusing to
have two methods to tweak slow-start behavior, one is for auto-reduce
parallelism and the other one is for all other cases.
https://github.com/apache/tez/blob/73bcabd2bca2536bf4f3673443a8dcdaaf79a4eb/tez-dag/src/main/java/org/apache/tez/dag/app/dag/impl/VertexImpl.java#L2893-L2897
My ideas are
- use `TEZ_SHUFFLE_VERTEX_MANAGER_{MIN, MAX}_SRC_FRACTION`
- add params for auto parallelism, like `TEZ_AUTO_REDUCER_PARALLELISM_{MIN,
MAX}_SRC_FRACTION`, and use them only for auto parallelism
- add params to configure slow-start behavior, like `TEZ_SLOW_START_{MIN,
MAX}_SRC_FRACTION`, and use it for all cases, meaning Hive on Tez ignores
`TEZ_SHUFFLE_VERTEX_MANAGER_{MIN, MAX}_SRC_FRACTION` configured by a user
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 548959)
Time Spent: 0.5h (was: 20m)
> Make the slow-start behavior tunable
> ------------------------------------
>
> Key: HIVE-24485
> URL: https://issues.apache.org/jira/browse/HIVE-24485
> Project: Hive
> Issue Type: Improvement
> Components: Hive, Tez
> Affects Versions: 3.1.2, 4.0.0
> Reporter: okumin
> Assignee: okumin
> Priority: Major
> Labels: pull-request-available
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> This ticket would enable users to configure the timing of slow-start with
> `tez.shuffle-vertex-manager.min-src-fraction` and
> `tez.shuffle-vertex-manager.max-src-fraction`.
> Hive on Tez currently doesn't honor these parameters and ShuffleVertexManager
> always uses the default value.
> We can control the timing to start vertexes the accuracy of estimated input
> size if we can tweak these ones. This is useful when a vertex has tasks that
> process a different amount of data.
>
> We can reproduce the issue with this query.
> {code:java}
> SET hive.tez.auto.reducer.parallelism=true;
> SET hive.tez.min.partition.factor=1.0; -- enforce auto-parallelism
> SET tez.shuffle-vertex-manager.min-src-fraction=0.55;
> SET tez.shuffle-vertex-manager.max-src-fraction=0.95;
> CREATE TABLE mofu (name string);
> INSERT INTO mofu (name) VALUES ('12345');
> SELECT name, count(*) FROM mofu GROUP BY name;{code}
> The fractions are ignored.
> {code:java}
> 2020-12-04 11:41:42,484 [INFO] [Dispatcher thread {Central}]
> |vertexmanager.ShuffleVertexManagerBase|: Settings minFrac: 0.25 maxFrac:
> 0.75 auto: true desiredTaskIput: 256000000
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
