[BEAM-10774] GBK Python streaming load tests are too slow

• Type: Bug
• Status: Open
• Priority: P3
• Resolution: Unresolved
• Component/s: runner-dataflow, sdk-py-core, testing
• Assignee: Unassigned
• Reporter: Kamil Wasilewski
• Created: 20/Aug/20 09:52
• Updated: 04/Jun/22 17:13

Description

The following GBK streaming test cases take too long on Dataflow:

 

1) 2GB of 10B records

2) 2GB of 100B records

4) fanout 4 times with 2GB 10-byte records total

5) fanout 8 times with 2GB 10-byte records total

 

Each of them needs at least 1 hour to execute, which is way too long for one Jenkins job.

Job's definition: https://github.com/apache/beam/blob/master/.test-infra/jenkins/job_LoadTests_GBK_Python.groovy

Test pipeline: https://github.com/apache/beam/blob/master/sdks/python/apache_beam/testing/load_tests/group_by_key_test.py

It is probable that those cases are too extreme. The first two cases involve grouping 20M unique keys, which is a stressful operation. A solution might be to overhaul the cases so that they would be less complex.

Both the current production Dataflow runner and the new Dataflow Runner V2 were tested.

Issue Links

• is related to: BEAM-10675 Create Streaming GBK Python Load Test Jenkins Job