Code here benchmarks message delivery latency with various services listed below. This tool uses Go and the official upstream client libraries.
Install gcloud. Login with a Google Cloud Platform account that has IAM permissions to list, create,
and delete Pub/Sub topics and subscriptions.
gcloud auth login --account [[email protected]]
gcloud auth application-default login --account [[email protected]]
Install some decently recent version of Golang. You might also need to install a package that provides
the gcc compiler. The build-essential package provides this for Ubuntu/Debian.
- Use separate machines for load generation and consumption.
- Run
scripts/./install.shto compile Go binaries. Default target OS isdarwin. You can change to others likelinux.
- Google designed Cloud Pub/Sub for high message volumes. Low volumes will cause higher latency. This has to do with how Pub/Sub balances the pull requests between listeners.
bin/pubsub/publisher \
-projectID warpspeed-testing \
-topicPrefix dxiatest \
-testLabelKey dxiatest
2021/07/29 23:28:40 created topicID: projects/warpspeed-testing/topics/dxiatest-1627601318562471531
2021/07/29 23:28:41 created subscriptionID: projects/warpspeed-testing/subscriptions/dxiatest-1627601318562471531
press enter to start publishing messages at 100 RPS...
[Before continuning, start the subscribers with the subscription ID above after the last slash. See below.]
2021/07/29 23:34:07 successfully published 8107 messages, failed 0
See more options with -h.
bin/kafka/./producer \
-brokerHost my-broker.com \
-brokerPort 9092 \
-topicPrefix dxiatest \
-messageSize 100 \
-RPS 1000
2021/08/05 20:36:44 created topic dxiatest-1628195803751335869
press enter to start producing 100 byte messages at 1000 RPS for 5m0s...
bin/subscriber \
-projectID warpspeed-testing \
-subscriptionID dxiatest-1627601318562471531
...
2021/07/29 23:39:02 got message 1627601942927737706 at 1627601942955833627, latency 28 ms
2021/07/29 23:39:03 burning in for 35 more seconds
2021/07/29 23:39:03 got message 1627601943118298744 at 1627601943143695090, latency 25 ms
2021/07/29 23:39:03 burning in for 35 more seconds
...
2021/07/29 23:41:12 got message 1627602072178202997 at 1627602072206435115, latency 28 ms
2021/07/29 23:41:12 got message 1627602072447232089 at 1627602072490150539, latency 43 ms
2021/07/29 23:41:12 got message 1627602072487471149 at 1627602072523003862, latency 36 ms
2021/07/29 23:41:12 got message 1627602072566998024 at 1627602072678355582, latency 111 ms
2021/07/29 23:41:12 got message 1627602072724468599 at 1627602072750649353, latency 26 ms
2021/07/29 23:41:12 got message 1627602072815071921 at 1627602072854578547, latency 40 ms
...
See more options with -h.
You can get latency quantiles while the subscriber is running.
watch -n 5 'curl --silent localhost:8080/metrics | grep pubsub'
Every 5.0s: curl localhost:8080/metrics | grep pubsub
pubsub_latency{quantile="0.5"} 26.640585
pubsub_latency{quantile="0.9"} 46.789227
pubsub_latency{quantile="0.99"} 470.668934
pubsub_latency_sum 68824.19071099993
pubsub_latency_count 1460
