Benchmark to test different kafka go clients to compare them under the same conditions.

This started as a proof of concept of creating some benchmarks that could run directly on github actions, publishing the results in a dashboard associated with the same repository.

This benchmark has been greatly inspired by Matt Howlett's benchmark. I shamelessly used what it has instead of reading the documentation of the different clients :S.

The clients tested in this benchmark are:

• Confluent-kafka-Go
• Sarama
• KafkaGo

In order to run the benchmarks we need the following cli tools installed:

• Ginkgo.

In order to run the benchmark tests we need a working kafka image. In this benchmark I've used docker-based Kafka instances but it's possible to use any kafka.

To run the benchmarks we use the following syntax:

ginkgo -focus=${type} ./... -- -test.bench=. -library=${client} -num=${num-messages} -size=${size-messages} -brokers=${brokers} -topic=${topic}

It can be one of the following:

• producer: To test the client producers against the kafka brokers selected.
• consumer: To test the client consumers against the kafka brokers selected.
• lag: To perform a lag check agains a topic to see if the producer sent all the messages as expected.

The syntax of the lag is slightly different to the other two types.

ginkgo -focus=lag ./... -- -test.bench=. -num=${num-messages} -brokers=${brokers} -topic=${topic}

Where num-messages is the number of messages we would expect to find in that topic.

It can be one of the following:

• confluent: (producer only). To test the Confluent-kafka-Go producer.
• confluent-poll: (consumer only). To test the Confluent-kafka-Go polling-based consumer.
• confluent-channel: (consumer only). To test the Confluent-kafka-Go channel-based (and deprecated) consumer.
• sarama: To test the Sarama producer/consumer.
• kafkago: To test the KafkaGo producer/consumer.

It's an integer specifying the number of messages to produce/consume for the benchmark.

It's an integer specifying the size of messages to use for the benchmark. In case of the consumers the Kafka instance will be prepopulated with messages of this size to be consumed.

Location of the kafka brokers to use for this benchmark. It includes the port. For instance localhost:9092.

The topic as string to use for this benchmark.

The output of the benchmarks is automatically display in the dashboard. The results are downloaded from the github actions, where are stored as artifacts, and used as sources when building the results dashboard.

• https://gguridi.github.io/benchmark-kafka-go-clients/

If a client doesn't appear in the graph is because the time used in the benchmark is so much higher than I didn't consider it necessary to perform further tests of it.

Special thanks to the components that I used for this:

• Matt Howlett's Benchmark
• Creative Tim Vue Material Dashboard