AWS Compute Blog

Scaling improvements when processing Apache Kafka with AWS Lambda

AWS Lambda is improving the automatic scaling behavior when processing data from Apache Kafka event-sources. Lambda is increasing the default number of initial consumers, improving how quickly consumers scale up, and helping to ensure that consumers don’t scale down too quickly. There is no additional action that you must take, and there is no additional cost.

Running Kafka on AWS

Apache Kafka is a popular open-source platform for building real-time streaming data pipelines and applications. You can deploy and manage your own Kafka solution on-premises or in the cloud on Amazon EC2.

Amazon Managed Streaming for Apache Kafka (MSK) is a fully managed service that makes it easier to build and run applications that use Kafka to process streaming data. MSK Serverless is a cluster type for Amazon MSK that allows you to run Kafka without having to manage and scale cluster capacity. It automatically provisions and scales capacity while managing the partitions in your topic, so you can stream data without thinking about right-sizing or scaling clusters. MSK Serverless offers a throughput-based pricing model, so you pay only for what you use. For more information, see Using Kafka to build your streaming application.

Using Lambda to consume records from Kafka

Processing streaming data can be complex in traditional, server-based architectures, especially if you must react in real-time. Many organizations spend significant time and cost managing and scaling their streaming platforms. In order to react fast, they must provision for peak capacity, which adds complexity.

Lambda and serverless architectures remove the undifferentiated heavy lifting when processing Kafka streams. You don’t have to manage infrastructure, can reduce operational overhead, lower costs, and scale on-demand. This helps you focus more on building streaming applications. You can write Lambda functions in a number of programming languages, which provide flexibility when processing streaming data.

Lambda event source mapping

Lambda can integrate natively with your Kafka environments as a consumer to process stream data as soon as it’s generated.

To consume streaming data from Kafka, you configure an event source mapping (ESM) on your Lambda functions. This is a resource managed by the Lambda service, which is separate from your function. It continually polls records from the topics in the Kafka cluster. The ESM optionally filters records and batches them into a payload. Then, it calls the Lambda Invoke API to deliver the payload to your Lambda function synchronously for processing.

As Lambda manages the pollers, you don’t need to manage a fleet of consumers across multiple teams. Each team can create and configure their own ESM with Lambda handling the polling.

AWS Lambda event source mapping

AWS Lambda event source mapping

For more information on using Lambda to process Kafka streams, see the learning guide.

Scaling and throughput

Kafka uses partitions to increase throughput and spread the load of records to all brokers in a cluster.

The Lambda event source mapping resource includes pollers and processors. Pollers have consumers that read records from Kafka partitions. The poller assigners send them to processors which batch the records and invoke your function.

When you create a Kafka event source mapping, Lambda allocates consumers to process all partitions in the Kafka topic. Previously, Lambda allocated a minimum of one processor for a consumer.

Lambda previous initial scaling

Lambda previous initial scaling

With these scaling improvements, Lambda allocates multiple processors to improve processing. This reduces the possibility of a single invoke slowing down the entire processing stream.

Lambda updated initial scaling

Lambda updated initial scaling

Each consumer sends records to multiple processors running in parallel to handle increased workloads. Records in each partition are only assigned to a single processor to maintain order.

Lambda automatically scales up or down the number of consumers and processors based on workload. Lambda samples the consumer offset lag of all the partitions in the topic every minute. If the lag is increasing, this means Lambda can’t keep up with processing the records from the partition.

The scaling algorithm accounts for the current offset lag, and also the rate of new messages added to the topic. Lambda can reach the maximum number of consumers within three minutes to lower the offset lag as quickly as possible. Lambda is also reducing the scale down behavior to ensure records are processed more quickly and latency is reduced, particularly for bursty workloads.

Total processors for all pollers can only scale up to the total number of partitions in the topic.

After successful invokes, the poller periodically commits offsets to the respective brokers.

Lambda further scaling

Lambda further scaling

You can monitor the throughput of your Kafka topic using consumer metrics consumer_lag and consumer_offset.

To check how many function invocations occur in parallel, you can also monitor the concurrency metrics for your function. The concurrency is approximately equal to the total number of processors across all pollers, depending on processor activity. For example, if three pollers have five processors running for a given ESM, the function concurrency would be approximately 15 (5 + 5 + 5).

Seeing the improved scaling in action

There are a number of Serverless Patterns that you can use to process Kafka streams using Lambda. To set up Amazon MSK Serverless, follow the instructions in the GitHub repo:

  1. Create an example Amazon MSK Serverless topic with 1000 partitions.
  2. ./ --create --bootstrap-server "{bootstrap-server}" --command-config --replication-factor 3 --partitions 1000 --topic msk-1000p
  3. Add records to the topic using a UUID as a key to distribute records evenly across partitions. This example adds 13 million records.
  4. for x in {1..13000000}; do echo $(uuidgen -r),message_$x; done | ./ --broker-list "{bootstrap-server}" --topic msk-1000p --producer.config --property parse.key=true --property key.separator=, --producer-property acks=all
  5. Create a Python function based on this pattern, which logs the processed records.
  6. Amend the function code to insert a delay of 0.1 seconds to simulate record processing.
  7. import json
    import base64
    import time
    def lambda_handler(event, context):
        # Define a variable to keep track of the number of the message in the batch of messages
        # Looping through the map for each key (combination of topic and partition)
        for record in event['records']:
            for messages in event['records'][record]:
                print("Record number: " + str(i))
                print("Topic: " + str(messages['topic']))
                print("Partition: " + str(messages['partition']))
                print("Offset: " + str(messages['offset']))
                print("Timestamp: " + str(messages['timestamp']))
                print("TimestampType: " + str(messages['timestampType']))
                if None is not messages.get('key'):
                if None is not messages.get('value'):
                print("Key = " + str(decodedKey))
                print("Value = " + str(decodedValue))
        return {
            'statusCode': 200,
  8. Configure the ESM to point to the previously created cluster and topic.
  9. Use the default batch size of 100. Set the StartingPosition to TRIM_HORIZON to process from the beginning of the stream.
  10. Deploy the function, which also adds and configures the ESM.
  11. View the Amazon CloudWatch ConcurrentExecutions and OffsetLag metrics to view the processing.

With the scaling improvements, once the ESM is configured, the ESM and function scale up to handle the number of partitions.

Lambda automatic scaling improvement graph

Lambda automatic scaling improvement graph

Increasing data processing throughput

It is important that your function can keep pace with the rate of traffic. A growing offset lag means that the function processing cannot keep up. If the age is high in relation to the stream’s retention period, you can lose data as records expire from the stream.

This value should generally not exceed 50% of the stream’s retention period. When the value reaches 100% of the stream retention period, data is lost. One temporary solution is to increase the retention time of the stream. This gives you more time to resolve the issue before losing data.

There are several ways to improve processing throughput.

  1. Avoid processing unnecessary records by using content filtering to control which records Lambda sends to your function. This helps reduce traffic to your function, simplifies code, and reduces overall cost.
  2. Lambda allocates processors across all pollers based on the number of partitions up to a maximum of one concurrent Lambda function per partition. You can increase the number of processing Lambda functions by increasing the number of partitions.
  3. For compute intensive functions, you can increase the memory allocated to your function, which also increases the amount of virtual CPU available. This can help reduce the duration of a processing function.
  4. Lambda polls Kafka with a configurable batch size of records. You can increase the batch size to process more records in a single invocation. This can improve processing time and reduce costs, particularly if your function has an increased init time. A larger batch size increases the latency to process the first record in the batch, but potentially decreases the latency to process the last record in the batch. There is a tradeoff between cost and latency when optimizing a partition’s capacity and the decision depends on the needs of your workload.
  5. Ensure that your producers evenly distribute records across partitions using an effective partition key strategy. A workload is unbalanced when a single key dominates other keys, creating a hot partition, which impacts throughput.

See Increasing data processing throughput for some additional guidance.


Today, AWS Lambda is improving the automatic scaling behavior when processing data from Apache Kafka event-sources. Lambda is increasing the default number of initial consumers, improving how quickly they scale up, and ensuring they don’t scale down too quickly. There is no additional action that you must take, and there is no additional cost.

You can explore the scaling improvements with your existing workloads or deploy an Amazon MSK cluster and try one of the patterns to measure processing time.

To explore using Lambda to process Kafka streams, see the learning guide.

For more serverless learning resources, visit Serverless Land.