[SPARK-33635] Performance regression in Kafka read - ASF JIRA

Rank to Top

Rank to Bottom

Attach files

Attach Screenshot

Bulk Copy Attachments

Bulk Move Attachments

Voters

Watch issue

Watchers

Create sub-task

Convert to sub-task

Link

Clone

Labels

Update Comment Author

Replace String in Comment

Update Comment Visibility

Delete Comments

XML

Word

Printable

JSON

Details

Type: Bug
Status: Resolved
Priority: Blocker
Resolution: Fixed
Affects Version/s: 3.0.0, 3.0.1
Fix Version/s: 3.0.2, 3.1.1
Component/s: SQL
Labels:
None
Environment:

Hide

A simple 5 node system. A simple data row of csv data in kafka, evenly distributed between the partitions.

Open JDK 1.8.0.252

Spark in stand alone - 5 nodes, 10 workers (2 worker per node, each locked to a distinct NUMA group)
kafka (v 2.3.1) cluster - 5 nodes (1 broker per node).
Centos 7.7.1908

1 topic, 10 partiions, 1 hour queue life

(this is just one of clusters we have, I have tested on all of them and theyall exhibit the same performance degredation)

Show
A simple 5 node system. A simple data row of csv data in kafka, evenly distributed between the partitions. Open JDK 1.8.0.252 Spark in stand alone - 5 nodes, 10 workers (2 worker per node, each locked to a distinct NUMA group) kafka (v 2.3.1) cluster - 5 nodes (1 broker per node). Centos 7.7.1908 1 topic, 10 partiions, 1 hour queue life (this is just one of clusters we have, I have tested on all of them and theyall exhibit the same performance degredation)

Target Version/s:

3.0.2, 3.1.0

Description

I have observed a slowdown in the reading of data from kafka on all of our systems when migrating from spark 2.4.5 to Spark 3.0.0 (and Spark 3.0.1)

I have created a sample project to isolate the problem as much as possible, with just a read all data from a kafka topic (see https://github.com/codegorillauk/spark-kafka-read ).

With 2.4.5, across multiple runs,
I get a stable read rate of 1,120,000 (1.12 mill) rows per second

With 3.0.0 or 3.0.1, across multiple runs,
I get a stable read rate of 632,000 (0.632 mil) rows per second

The represents a 44% loss in performance. Which is, a lot.

I have been working though the spark-sql-kafka-0-10 code base, but change for spark 3 have been ongoing for over a year and its difficult to pin point an exact change or reason for the degradation.

I am happy to help fix this problem, but will need some assitance as I am unfamiliar with the spark-sql-kafka-0-10 project.

A sample of the data my test reads (note: its not parsing csv - this is just test data)
1606921800000,001e0610e532,lightsense,tsl250rd,intensity,21853,53.262,acceleration_z,651,ep,290,commit,913,pressure,138,pm1,799,uv_intensity,823,idletime,-372,count,-72,ir_intensity,185,concentration,-61,flags,-532,tx,694.36,ep_heatsink,-556.92,acceleration_x,-221.40,fw,910.53,sample_flow_rate,-959.60,uptime,-515.15,pm10,-768.03,powersupply,214.72,magnetic_field_y,-616.04,alphasense,606.73,AoT_Chicago,053,Racine Ave & 18th St Chicago IL,41.857959,-87.65642700000002,AoT Chicago (S) [C],2017/12/15 00:00:00,