Introduce Default Replication Worker Performance Test Harness #20956

davinchia · 2023-01-03T01:11:41Z

What

Introduce a performance test harness for the default replication worker to make it easy for devs to test effect of changes on platform throughput.

The current set up is designed to be run manually. In the future, we can look into integrating this report into our build pipelines. For now, this is good enough as I wanted to start somewhere.

How

Set up a source that spits out dynamic messages.
Set up a destination that does nothing to mimic the fastest destination.
Set up the ReplicationPerformanceClass with an E2E sync that has performs all the same steps as the current DefaultReplicationWorker.
Add Java Micro Benchmark as a simple test harness.

The general idea is to use JMH to run the test n number of times (currently 4 times). The dev can then look at logs to see throughput and how it varies.

As of this PR, we see general platform throughput of ~ 20 - 25 MB/s.

🚨 User Impact 🚨

Are there any breaking changes? What is the end result perceived by the user? If yes, please merge this PR with the 🚨🚨 emoji so changelog authors can further highlight this if needed.

Pre-merge Checklist

Expand the relevant checklist and delete the others.

New Connector

Community member or Airbyter

Airbyter

If this is a community PR, the Airbyte engineer reviewing this PR is responsible for the below items.

Create a non-forked branch based on this PR and test the below items on it
Build is successful
If new credentials are required for use in CI, add them to GSM. Instructions.
/test connector=connectors/<name> command is passing
New Connector version released on Dockerhub by running the /publish command described here
After the connector is published, connector added to connector index as described here
Seed specs have been re-generated by building the platform and committing the changes to the seed spec files, as described here

Updating a connector

Community member or Airbyter

Grant edit access to maintainers (instructions)
Secrets in the connector's spec are annotated with airbyte_secret
Unit & integration tests added and passing. Community members, please provide proof of success locally e.g: screenshot or copy-paste unit, integration, and acceptance test output. To run acceptance tests for a Python connector, follow instructions in the README. For java connectors run ./gradlew :airbyte-integrations:connectors:<name>:integrationTest.
Code reviews completed
Documentation updated
- Connector's README.md
- Connector's bootstrap.md. See description and examples
- Changelog updated in docs/integrations/<source or destination>/<name>.md including changelog. See changelog example
PR name follows PR naming conventions

Airbyter

If this is a community PR, the Airbyte engineer reviewing this PR is responsible for the below items.

Create a non-forked branch based on this PR and test the below items on it
Build is successful
If new credentials are required for use in CI, add them to GSM. Instructions.
/test connector=connectors/<name> command is passing
New Connector version released on Dockerhub and connector version bumped by running the /publish command described here

Connector Generator

Issue acceptance criteria met
PR name follows PR naming conventions
If adding a new generator, add it to the list of scaffold modules being tested
The generator test modules (all connectors with -scaffold in their name) have been updated with the latest scaffold by running ./gradlew :airbyte-integrations:connector-templates:generator:testScaffoldTemplates then checking in your changes
Documentation which references the generator is updated as needed

Tests

Unit

Put your unit tests output here.

Integration

Put your integration tests output here.

Acceptance

Put your acceptance tests output here.

evantahler · 2023-01-03T03:25:15Z

...ommons-worker/src/test/java/io/airbyte/workers/general/ReplicationWorkerPerformanceTest.java

+    workerThread.start();
+    workerThread.join();


Is this test in the same process, but using different threads? If that's the case, then this also bypasses the stdio streams as well, which also present their own bottlenecks (cc @colesnodgrass)

Agree. Let me see if I can quickly work something up.

It does show that the core serialising isn't the bottleneck like we thought.

airbyte-commons-worker/src/test/java/io/airbyte/workers/general/StubAirbyteMapper.java

airbyte-commons-worker/src/test/java/io/airbyte/workers/general/LimitedAirbyteSource.java

…tructured-concurrency

davinchia · 2023-01-04T22:20:05Z

...ommons-worker/src/test/java/io/airbyte/workers/general/ReplicationWorkerPerformanceTest.java

+  @Benchmark
+  // SampleTime = the time taken to run the benchmarked method. Use this because we only care about
+  // the time taken to sync the entire dataset.
+  @BenchmarkMode(Mode.SampleTime)


@colesnodgrass unfortunately because these annotations don't have the target annotation type, we cannot fold them into a meta annotation. I think this is fine for now.

davinchia · 2023-01-04T22:20:36Z

fyi @cgardens

davinchia · 2023-01-04T22:27:11Z

Sample logging output:

      "recordsCommitted" : 1000000
    }
  } ]
}
2023-01-04 22:25:31 INFO i.a.w.g.DefaultReplicationWorker(getReplicationOutput):436 - failures: [ ]
2023-01-04 22:25:31 INFO i.a.w.g.ReplicationWorkerPerformanceTest(executeOneSync):96 - MBs read: 178, Time taken sec: 8.312, MB/s: 21.414821944177096
2023-01-04 22:25:31 INFO i.a.c.i.LineGobbler(voidCall):114 - 
2023-01-04 22:25:31 INFO i.a.c.i.LineGobbler(voidCall):114 - ----- END REPLICATION -----
2023-01-04 22:25:31 INFO i.a.c.i.LineGobbler(voidCall):114 - 
8.493 s/op
                 executeOneSync·p0.00:   8.313 s/op
                 executeOneSync·p0.50:   8.493 s/op
                 executeOneSync·p0.90:   8.674 s/op
                 executeOneSync·p0.95:   8.674 s/op
                 executeOneSync·p0.99:   8.674 s/op
                 executeOneSync·p0.999:  8.674 s/op
                 executeOneSync·p0.9999: 8.674 s/op
                 executeOneSync·p1.00:   8.674 s/op



Result "io.airbyte.workers.general.ReplicationWorkerPerformanceTest.executeOneSync":
  N = 4
  mean =      8.831 ±(99.9%) 4.006 s/op

  Histogram, s/op:
    [ 8.000,  8.125) = 0 
    [ 8.125,  8.250) = 0 
    [ 8.250,  8.375) = 1 
    [ 8.375,  8.500) = 0 
    [ 8.500,  8.625) = 1 
    [ 8.625,  8.750) = 1 
    [ 8.750,  8.875) = 0 
    [ 8.875,  9.000) = 0 
    [ 9.000,  9.125) = 0 
    [ 9.125,  9.250) = 0 
    [ 9.250,  9.375) = 0 
    [ 9.375,  9.500) = 0 
    [ 9.500,  9.625) = 0 
    [ 9.625,  9.750) = 1 
    [ 9.750,  9.875) = 0 

  Percentiles, s/op:
      p(0.0000) =      8.313 s/op
     p(50.0000) =      8.640 s/op
     p(90.0000) =      9.731 s/op
     p(95.0000) =      9.731 s/op
     p(99.0000) =      9.731 s/op
     p(99.9000) =      9.731 s/op
     p(99.9900) =      9.731 s/op
     p(99.9990) =      9.731 s/op
     p(99.9999) =      9.731 s/op
    p(100.0000) =      9.731 s/op


# Run complete. Total time: 00:00:36

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

NOTE: Current JVM experimentally supports Compiler Blackholes, and they are in use. Please exercise
extra caution when trusting the results, look into the generated code to check the benchmark still
works, and factor in a small probability of new VM bugs. Additionally, while comparisons between
different JVMs are already problematic, the performance difference caused by different Blackhole
modes can be very significant. Please make sure you use the consistent Blackhole mode for comparisons.

Benchmark                                                                 Mode  Cnt  Score   Error  Units
ReplicationWorkerPerformanceTest.executeOneSync                         sample    4  8.831 ± 4.006   s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.00    sample       8.313           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.50    sample       8.640           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.90    sample       9.731           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.95    sample       9.731           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.99    sample       9.731           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.999   sample       9.731           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p0.9999  sample       9.731           s/op
ReplicationWorkerPerformanceTest.executeOneSync:executeOneSync·p1.00    sample       9.731           s/op

davinchia · 2023-01-04T23:41:20Z

PTAL @colesnodgrass @jdpgrailsdev @evantahler @pmossman

Cole/Jonathan, one approval for either of you two is enough. Mainly tagging for awareness. Parker, same thing here. Want to make sure both platform teams are aware this is available.

Evan - you are already on the PR. Thought to tag you since you are also interested in this work!

Checkpoint: Use Java 19 features to test orchestrator speed.

b276b57

octavia-squidington-iv added area/platform issues related to the platform area/server labels Jan 3, 2023

Remove virtual threads.

097ed19

davinchia temporarily deployed to more-secrets January 3, 2023 01:22 — with GitHub Actions Inactive

davinchia temporarily deployed to more-secrets January 3, 2023 01:23 — with GitHub Actions Inactive

evantahler reviewed Jan 3, 2023

View reviewed changes

Merge remote-tracking branch 'origin/master' into davinchia/java-19-s…

4ee048f

…tructured-concurrency

davinchia temporarily deployed to more-secrets January 4, 2023 02:17 — with GitHub Actions Inactive

Revert Java 19 changes. Get things working again.

bec9d04

davinchia temporarily deployed to more-secrets January 4, 2023 03:20 — with GitHub Actions Inactive

davinchia added 2 commits January 3, 2023 19:22

Remove testing changes.

a05ab50

Format.

c508e43

octavia-squidington-iv removed area/platform issues related to the platform area/server labels Jan 4, 2023

davinchia temporarily deployed to more-secrets January 4, 2023 03:25 — with GitHub Actions Inactive

davinchia temporarily deployed to more-secrets January 4, 2023 03:26 — with GitHub Actions Inactive

Add dependencies for JMH testing.

799eaaa

davinchia temporarily deployed to more-secrets January 4, 2023 05:09 — with GitHub Actions Inactive

Checkpoint: JMH setup and run. However not 100% sure this works now.

58f50c7

davinchia temporarily deployed to more-secrets January 4, 2023 05:39 — with GitHub Actions Inactive

davinchia temporarily deployed to more-secrets January 4, 2023 05:40 — with GitHub Actions Inactive

Test custom annotations.

46a03fa

davinchia temporarily deployed to more-secrets January 4, 2023 19:35 — with GitHub Actions Inactive

Merge branch 'master' into davinchia/java-19-structured-concurrency

3c3086f

davinchia temporarily deployed to more-secrets January 4, 2023 20:59 — with GitHub Actions Inactive

davinchia added 2 commits January 4, 2023 14:17

Move to dep.toml.

6b4c27e

Merge branch 'master' into davinchia/java-19-structured-concurrency

a09c219

davinchia marked this pull request as ready for review January 4, 2023 22:19

davinchia requested a review from a team as a code owner January 4, 2023 22:19

davinchia commented Jan 4, 2023

View reviewed changes

davinchia requested review from colesnodgrass, jdpgrailsdev and pmossman January 4, 2023 22:20

davinchia temporarily deployed to more-secrets January 4, 2023 22:21 — with GitHub Actions Inactive

Merge branch 'master' into davinchia/java-19-structured-concurrency

b1ac9a9

davinchia temporarily deployed to more-secrets January 4, 2023 22:42 — with GitHub Actions Inactive

davinchia added 4 commits January 4, 2023 15:28

Clean up testing comments.

5cbb99a

Better comments.

89e60b5

Fix PMD.

ce90cc2

Ignore JMH generated code.

7050698

davinchia temporarily deployed to more-secrets January 4, 2023 23:31 — with GitHub Actions Inactive

colesnodgrass approved these changes Jan 5, 2023

View reviewed changes

davinchia merged commit eeedd64 into master Jan 5, 2023

davinchia deleted the davinchia/java-19-structured-concurrency branch January 5, 2023 22:22

This was referenced Jan 6, 2023

Bump Airbyte version from 0.40.26 to 0.40.27 #21092

Closed

Bump Airbyte version from 0.40.26 to 0.40.27 #21121

Closed

github-actions bot mentioned this pull request Jan 6, 2023

Bump helm chart version reference to 0.43.15 #21125

Closed

This was referenced Jan 6, 2023

Bump Airbyte version from 0.40.26 to 0.40.27 #21130

Closed

Bump Airbyte version from 0.40.26 to 0.40.27 #21135

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Introduce Default Replication Worker Performance Test Harness #20956

Introduce Default Replication Worker Performance Test Harness #20956

davinchia commented Jan 3, 2023 •

edited

Loading

evantahler Jan 3, 2023

davinchia Jan 4, 2023

davinchia Jan 4, 2023

davinchia Jan 4, 2023

davinchia commented Jan 4, 2023

davinchia commented Jan 4, 2023

davinchia commented Jan 4, 2023

Introduce Default Replication Worker Performance Test Harness #20956

Introduce Default Replication Worker Performance Test Harness #20956

Conversation

davinchia commented Jan 3, 2023 • edited Loading

What

How

Recommended reading order

🚨 User Impact 🚨

Pre-merge Checklist

Community member or Airbyter

Airbyter

Community member or Airbyter

Airbyter

Tests

evantahler Jan 3, 2023

Choose a reason for hiding this comment

davinchia Jan 4, 2023

Choose a reason for hiding this comment

davinchia Jan 4, 2023

Choose a reason for hiding this comment

davinchia Jan 4, 2023

Choose a reason for hiding this comment

davinchia commented Jan 4, 2023

davinchia commented Jan 4, 2023

davinchia commented Jan 4, 2023

davinchia commented Jan 3, 2023 •

edited

Loading