[IMPALA-6567] Dataload performance regression due to slow invalidate metadata

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.12.0
• Fix Version/s: Impala 2.12.0
• Component/s: Frontend
• Target Version: Impala 2.12.0
• Epic Color: ghx-label-7
• Assignee: Alexander Behm
• Reporter: Joe McDonnell
• Created: 23/Feb/18 01:13
• Updated: 07/Mar/18 17:12
• Resolved: 23/Feb/18 21:17

Description

Recent GVO builds intermittently have a functional dataload of almost 2 hours when it used to be ~30-35 minutes:
**

02:12:15 Loading TPC-DS data (logging to /home/ubuntu/Impala/logs/data_loading/load-tpcds.log)...
02:34:27 Loading workload 'tpch' using exploration strategy 'core' OK (Took: 22 min 12 sec)
02:34:35 Loading workload 'tpcds' using exploration strategy 'core' OK (Took: 22 min 20 sec)
04:11:40 Loading workload 'functional-query' using exploration strategy 'exhaustive' OK (Took: 119 min 25 sec)

 

This has happened on multiple runs (including some in progress):

https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/1370/

https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/1382/

https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/1383/ (missing some logs due to abort)

https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/1384/ (in progress)

https://jenkins.impala.io/job/ubuntu-16.04-from-scratch/1385/ (in progress)

 

Dataload creates a SQL script that invalidates each table created using an "invalidate metadata ${tablename}" command. There are 830 "invalidate metadata ${tablename}" calls in the invocation of this script (see IMPALA-6386 for why we do invalidate at the table level). Even so, this script should execute very quickly.

The impalad.INFO from the 1370 run shows that this script is taking a long time. The first invalidate metadata for functional tables is at 2:41 and the last invalidate metadata for this run of the invalidate script is at 3:17. 

The invalidate script runs twice. The second run begins at 3:19 and finishes at 4:11. 

 

Issue Links

• is broken by: IMPALA-5152 Frontend requests metadata for one table at a time in the query