Get Started Free
course: Apache Kafka® Internal Architecture

Hands On: Tuning the Apache Kafka Producer Client

8 min
Danica Fine

Danica Fine

Senior Developer Advocate (Presenter)

Hands On: Tuning the Apache Kafka Producer Client

Now that you have a better idea of Kafka internals at a high level, let’s dive into some performance tuning! In this exercise, we’ll be producing events to a Kafka topic in a Confluent Cloud cluster, tweaking client settings, and observing the impact they have on event throughput and latency.

Before we get into the nitty gritty, there are a couple of environment preparation steps that you’ll have to complete.

Exercise Setup

Complete the following steps to set up the environment used for this exercise. Prior to doing so, you will need to sign up for Confluent Cloud.

Let’s start with creating the cluster that we use for this exercise.

  1. Open URL https://confluent.cloud and log in to the Confluent Cloud console.
  2. Navigate to the default environment.

NOTE: If the Get started with Confluent Cloud tutorial appears in the right side of the console, click Leave tutorial.

  1. Create a basic cluster named perf-test-cluster.

NOTE: If multiple people within your organization will be going through this exercise, append a unique string, e.g., &ltlast name&gt, to the cluster name to prevent your exercise from conflicting with others.

Now that we have our cluster, let’s create the topic that we will be producing events to.

  1. Click Topics.
  2. Click Create topic.
  3. Enter perf-test-topic in the Topic name field.
  4. Enter 4 in the Number of partitions field.
  5. Click Create with defaults.

We also need to create a client configuration file for the cluster that will be needed by the kafka-producer-perf-test command during the exercise.

  1. Click Cluster overview.
  2. Click Configure a client.

Since this command utilizes an embedded Java producer to write events to the topic, we need to click the corresponding Java client type.

  1. Click Java.

Here we see the client connection configs which we can copy using the provided button. As you can see, one of the parameters needs to be updated so that it includes a valid cluster API key and secret value. We can create these using the option that is conveniently available on this page.

  1. Click Create Kafka cluster API key.

As you can now see, clicking this option opens a new window that contains the new API key and secret.

  1. Assign a name of perf-test-cluster.
  2. Click the Continue button.

After creating the API key and secret, notice the client connection configs have been updated to include these values. We can now copy these configs and use them to create the needed client configuration file on our Java client machine.

  1. Click the Copy button.
  2. Minimize the Confluent Cloud console.

We will now create the local client configuration file containing the Confluent Cloud connection settings.

  1. Open a terminal window.
  2. Run command nano java.config.
  3. Right-click the nano edit window and click Paste.

The sample client configuration settings include properties that are needed if the client is using Schema Registry as well as a couple of other properties that are needed by earlier versions of the Java client. We won’t be using these properties so we will delete them.

  1. Delete all lines that follow sasl.mechanism=PLAIN.
  2. Save and close java.config.

We will use kafka-producer-perf-test.sh and the confluent CLI during the exercise. Downloading Confluent Platform Community components includes both of these.

  1. Download Confluent Platform:

  2. Run the following command to add the Confluent Platform bin directory to your path:

    echo 'export PATH=$PATH:/home/training/confluent-<version>/bin/' >> ~/.bashrc
  3. To update the Confluent CLI to its latest version, run the following command and complete the prompts that appear:

    confluent update --major

This concludes the exercise setup.

Exercise Steps

We’ll be using a performance test script kafka-producer-perf-test.sh to test different producer configuration settings to see how they impact event throughput and latency. Before starting, let’s be sure the client configuration file that will be used by our performance test script to establish a connection with the cluster is available. This file contains both the cluster endpoint as well as authentication settings, so it’s pretty important to have it set up properly. Let’s check out the config file and make sure all the right details are in it.

  1. Run command:

    cat java.config

You should see the cluster endpoint as well as your authentication settings. If your config file doesn’t quite look like what you saw in the exercise video, revisit the exercise setup instructions before continuing.

If everything looks good, we can now begin testing the producer client performance. To start off, we’ll just use the default producer configuration values. Two of the most important producer settings related to event throughput and latency are linger.ms and batch.size. The linger setting is the maximum amount of time that the producer will wait while adding events to a record batch before it’s flushed. The batch size setting, of course, is the maximum size that a batch can be before it’s flushed. The defaults for these settings are 0 and about 16 kilobytes, respectively. In the next few tests, we’ll alter the values for these settings and observe how they impact the average throughput and latency.

A quick thing to note before we start running our tests is that we’ll be throttling the record output of the script to make it easier to observe throughput and latency variations.

With that, let’s kick off a test with the default settings.

You’ll notice a couple key options in the performance test command. We’ll set the record size to 1000 bytes and specify that the test run through 3000 records with a throughput of 200. But, of course, the most important property is producer-props where we set linger and batch size. We’ll use the --print-metrics flag to output metrics, limiting the output using grep.

  1. Run command:

    kafka-producer-perf-test \
        --producer.config /home/training/java.config \
        --throughput 200 \
        --record-size 1000 \
        --num-records 3000 \
        --topic perf-test-topic \
        --producer-props linger.ms=0 batch.size=16384 \
        --print-metrics | grep \
    "3000 records sent\|\
    producer-metrics:outgoing-byte-rate\|\
    producer-metrics:bufferpool-wait-ratio\|\
    producer-metrics:record-queue-time-avg\|\
    producer-metrics:request-latency-avg\|\
    producer-metrics:batch-size-avg"

You should see output similar to the following:

Metriclinger.ms 0 batch.size 16384linger.ms 100 batch.size 16384linger.ms 100 batch.size 300000linger.ms 1500 batch.size 300000
batch-size-avg1215.030
bufferpool-wait-ratio0.000
outgoing-byte-rate75594.490
record-queue-time-avg4.229
request-latency-avg43.292

The output you see could be a bit different than shown here, but you should see something similar. In this case, the average batch size was 1215, much lower than the default batch size. This is easily explained by the linger value which was set to 0. Since linger was so low, the batch is being flushed almost as soon as the first record is added. Doing a bit of math here, since our record size equals 1000, it appears that one record is being added to each batch before linger triggers the flush.

Let’s quickly review the remaining output values. The bufferpool wait ratio 0 value indicates that batches are never waiting on previously sent requests in order to be flushed, meaning that Confluent Cloud brokers are able to process requests as fast as we are sending them. Pretty cool!

Outgoing byte rate and request latency average shows the throughput and latency results. And finally we have record queue time, showing how long that records are remaining in the batch prior to being flushed.

Now that we have a good understanding of the output metrics, let’s run the same test, but this time we’ll increase linger from 0 to 100 ms, giving records more time to be added to a batch before it’s flushed.

  1. Run command:

    kafka-producer-perf-test \
        --producer.config /home/training/java.config \
        --throughput 200 \
        --record-size 1000 \
        --num-records 3000 \
        --topic perf-test-topic \
        --producer-props linger.ms=100 batch.size=16384 \
        --print-metrics | grep \
    "3000 records sent\|\
    producer-metrics:outgoing-byte-rate\|\
    producer-metrics:bufferpool-wait-ratio\|\
    producer-metrics:record-queue-time-avg\|\
    producer-metrics:request-latency-avg\|\
    producer-metrics:batch-size-avg"

You should see output similar to the following:

Metriclinger.ms 0 batch.size 16384linger.ms 100 batch.size 16384linger.ms 100 batch.size 300000linger.ms 1500 batch.size 300000
batch-size-avg1215.03016164.670
bufferpool-wait-ratio0.0000.000
outgoing-byte-rate75594.49068866.596
record-queue-time-avg4.22995.457
request-latency-avg43.29253.649

Let’s analyze these results. With linger set to 100 ms, you’ll first notice that the average batch size is much higher—it’s actually just lower than the batch size we specified in the properties. This indicates that batch size is what’s triggering the batch to be flushed now. Notice also that throughput has decreased and latency has increased compared to the previous test. That’s actually not what we want.

Let’s rerun the test, increasing batch size to 300000.

  1. Run command:

    kafka-producer-perf-test \
        --producer.config /home/training/java.config \
        --throughput 200 \
        --record-size 1000 \
        --num-records 3000 \
        --topic perf-test-topic \
        --producer-props linger.ms=100 batch.size=300000 \
        --print-metrics | grep \
    "3000 records sent\|\
    producer-metrics:outgoing-byte-rate\|\
    producer-metrics:bufferpool-wait-ratio\|\
    producer-metrics:record-queue-time-avg\|\
    producer-metrics:request-latency-avg\|\
    producer-metrics:batch-size-avg"

You should see output similar to the following:

Metriclinger.ms 0 batch.size 16384linger.ms 100 batch.size 16384linger.ms 100 batch.size 300000linger.ms 1500 batch.size 300000
batch-size-avg1215.03016164.67023720.164
bufferpool-wait-ratio0.0000.0000.000
outgoing-byte-rate75594.49068866.59668720.642
record-queue-time-avg4.22995.457109.070
request-latency-avg43.29253.64963.836

Looking at the output results, we see that the average batch size has increased, but not by that much. This indicates that the linger setting isn’t allowing enough time for batches to reach the limit set by the batch size parameter. Throughput and latency are still moving in the wrong direction.

The next obvious thing to do is increase the linger time. Let’s see how this affects the output.

  1. Run command:

    kafka-producer-perf-test \
        --producer.config /home/training/java.config \
        --throughput 200 \
        --record-size 1000 \
        --num-records 3000 \
        --topic perf-test-topic \
        --producer-props linger.ms=1500 batch.size=300000 \
        --print-metrics | grep \
    "3000 records sent\|\
    producer-metrics:outgoing-byte-rate\|\
    producer-metrics:bufferpool-wait-ratio\|\
    producer-metrics:record-queue-time-avg\|\
    producer-metrics:request-latency-avg\|\
    producer-metrics:batch-size-avg"

You should see output similar to the following:

Metriclinger.ms 0 batch.size 16384linger.ms 100 batch.size 16384linger.ms 100 batch.size 300000linger.ms 1500 batch.size 300000
batch-size-avg1215.03016164.67023720.164275700.182
bufferpool-wait-ratio0.0000.0000.0000.000
outgoing-byte-rate75594.49068866.59668720.64268246.401
record-queue-time-avg4.22995.457109.0701406.273
request-latency-avg43.29253.64963.836226.091

With linger set to 1500 ms, we can see that the batch sizes are pretty close to the batch size limit. But we’re still losing ground in throughput and latency. It appears that the producer defaults for linger and batch size are actually the best choice for an application that’s producing 200 records per second of 1000 bytes each.

In this exercise, we got a pretty good feel for producer settings and saw how altering these configurations can affect throughput and latency. Although we only checked these settings for a specific application based on number of records and record size, we encourage you to continue playing around with these parameters and see how the producer configs impact an application. A couple of things you might try are:

  • Remove the throttle on the performance test by setting --throughput equal to -1
  • Increase the --num-records and --record-size values

Then observe the result, paying close attention to all metrics.

Note

When you’re finished running this exercise on your own, you’ll need to tear down the exercise environment, deleting the Confluent Cloud cluster to prevent it from unnecessarily accruing cost and exhausting your promotional credits.

From a terminal window, log into the CLI using the --save flag so that your credentials are saved for use in later commands.

  1. Run command:

    confluent login --save 
  2. Enter email and password.

List the environments available to you.

  1. Run command:

    confluent environment list

With the environment ID, now you list the available clusters.

  1. List the clusters in the environment and their IDs:

    confluent kafka cluster list \
        --environment <environment ID> 

We used the performance test cluster in this exercise, so we’ll use that ID along with the environment ID to run the delete cluster command.

  1. Delete the perf-test-cluster cluster:

    confluent kafka cluster delete <perf-test-cluster ID> \
        --environment <environment ID> 

As a sanity check, we’ll list the environment clusters again to confirm that the performance test cluster no longer exists.

  1. Confirm the perf-test-cluster cluster no longer exists in the environment:

    confluent kafka cluster list \
        --environment <environment ID> 

The cluster is not listed so the environment tear down is complete. And with that step complete, we’ve successfully wrapped up our first Kafka internals exercise! You should now have a pretty good idea of how producer configuration settings can impact your application and how to fine-tune these settings to optimize throughput and latency.

Use the promo code INTERNALS101 to get $25 of free Confluent Cloud usage

Disagree? If you believe that any of these rules do not necessarily support our goal of serving the Apache Kafka community, feel free to reach out to your direct community contact in the group or community@confluent.io

Be the first to get updates and new content

We will only share developer content and updates, including notifications when new content is added. We will never send you sales emails. 🙂 By subscribing, you understand we will process your personal information in accordance with our Privacy Statement.