Rajesh Balamohan created HIVE-24262:
---------------------------------------
Summary: Optimise NullScanTaskDispatcher for cloud storage
Key: HIVE-24262
URL: https://issues.apache.org/jira/browse/HIVE-24262
Project: Hive
Issue Type: Improvement
Reporter: Rajesh Balamohan
{noformat}
select count(DISTINCT ss_sold_date_sk) from store_sales;
----------------------------------------------------------------------------------------------
VERTICES MODE STATUS TOTAL COMPLETED RUNNING PENDING
FAILED KILLED
----------------------------------------------------------------------------------------------
Map 1 .......... container SUCCEEDED 1 1 0 0
0 0
Reducer 2 ...... container SUCCEEDED 1 1 0 0
0 0
----------------------------------------------------------------------------------------------
VERTICES: 02/02 [==========================>>] 100% ELAPSED TIME: 5.55 s
----------------------------------------------------------------------------------------------
INFO : Status: DAG finished successfully in 5.44 seconds
INFO :
INFO : Query Execution Summary
INFO :
----------------------------------------------------------------------------------------------
INFO : OPERATION DURATION
INFO :
----------------------------------------------------------------------------------------------
INFO : Compile Query 102.02s
INFO : Prepare Plan 0.51s
INFO : Get Query Coordinator (AM) 0.01s
INFO : Submit Plan 0.33s
INFO : Start DAG 0.56s
INFO : Run DAG 5.44s
INFO :
----------------------------------------------------------------------------------------------
{noformat}
Reason for this is that, it ends up doing "isEmptyPath" check for every
partition path and takes lot of time in compilation phase.
If the parent directory of all paths belong to the same path, we could just do
a recursive listing just once (instead of listing each directory one at a time
sequentially) in cloud storage systems.
https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/physical/NullScanTaskDispatcher.java#L158
https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/physical/NullScanTaskDispatcher.java#L121
https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/physical/NullScanTaskDispatcher.java#L101
With a temp hacky fix, it comes down to 2 seconds from 100+ seconds.
{noformat}
INFO : Dag name: select count(DISTINCT ss_sold_...store_sales (Stage-1)
INFO : Status: Running (Executing on YARN cluster with App id
application_1602500203747_0003)
----------------------------------------------------------------------------------------------
VERTICES MODE STATUS TOTAL COMPLETED RUNNING PENDING
FAILED KILLED
----------------------------------------------------------------------------------------------
Map 1 .......... container SUCCEEDED 1 1 0 0
0 0
Reducer 2 ...... container SUCCEEDED 1 1 0 0
0 0
----------------------------------------------------------------------------------------------
VERTICES: 02/02 [==========================>>] 100% ELAPSED TIME: 1.23 s
----------------------------------------------------------------------------------------------
INFO : Status: DAG finished successfully in 1.20 seconds
INFO :
INFO : Query Execution Summary
INFO :
----------------------------------------------------------------------------------------------
INFO : OPERATION DURATION
INFO :
----------------------------------------------------------------------------------------------
INFO : Compile Query 0.85s
INFO : Prepare Plan 0.17s
INFO : Get Query Coordinator (AM) 0.00s
INFO : Submit Plan 0.03s
INFO : Start DAG 0.03s
INFO : Run DAG 1.20s
INFO :
----------------------------------------------------------------------------------------------
{noformat}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
