AWS Compute Blog

Build a custom Java runtime for AWS Lambda

This post is written by Christian Müller, Principal AWS Solutions Architect and Maximilian Schellhorn, AWS Solutions Architect

When running applications on AWS Lambda, you have the option to use either one of the managed runtime versions that AWS provides or bring your own custom runtime. The following blog post provides a walkthrough of how you can create and optimize a custom runtime for Java based Lambda functions.

Builders might rely on customized or experimental runtime behavior when creating solutions in the cloud. The Java ecosystem fosters innovation and encourages experiments with the current six-month release schedule for the latest runtime versions.

However, Lambda focuses on providing stable long-term support (LTS) versions. The official Lambda runtimes are built around a combination of operating system, programming language, and software libraries that are subject to maintenance and security updates. For example, the Lambda runtime for Java supports the LTS versions Java 8 Corretto and Java 11 Corretto as of April 2022. The Java 17 Corretto version is pending. In addition, there is no provided runtime for non LTS versions like Java 15 Corretto, Java 16 Corretto, or Java 18 Corretto.

To use other language versions, Lambda allows you to create custom runtimes. Custom runtimes allow builders to provide and configure their own runtimes for running their application code. To enable communication between your custom runtime and Lambda, you can use the runtime interface client library in Java.

With the introduction of modular runtime images in Java 9 (JEP 220), it is possible to include only the Java runtime modules that your application depends on. This reduces the overall runtime size and increases performance, especially during cold-starts. In addition, there are other techniques in Java, like class data sharing and tiered compilation, which allow you to reduce the startup time of your application even further.

To combine those capabilities, this blog post provides an overview for creating and deploying a minified Java runtime on Lambda by using Java 18 Corretto. For step-by-step instructions and prerequisites, refer to the official GitHub example.

Overview of the example

In the following example, you build a custom runtime for a basic Java application that writes request headers to Amazon DynamoDB and is fronted by Amazon API Gateway.

Application architecture

The following diagram summarizes the steps to create the application and the custom runtime:

Steps to create the application custom runtime

  1. Download the preferred Java version and take advantage of jdeps, jlink and class data sharing to create a minified and optimized Java runtime based on the application code (function.jar).
  2. Create a bootstrap file with optimized starting instructions for the application.
  3. Package the application code, the optimized Java runtime, and the bootstrap file as a zip file.
  4. Deploy the runtime, including the app, to Lambda. For example, using the AWS Cloud Development Kit (CDK)

Steps 1–3 are automated and abstracted via Docker. The following section provides a high-level walkthrough of the build and deployment process. For the full version, see the Dockerfile in the GitHub example.

Creating the optimized Java runtime

1. Download the desired Java version and copy the local application code to the Docker environment and build it with Maven:

FROM amazonlinux:2

...

# Update packages and install Amazon Corretto 18, Maven and Zip
RUN yum -y update
RUN yum install -y java-18-amazon-corretto-devel maven zip

...

# Copy the software folder to the image and build the function
COPY software software
WORKDIR /software/example-function
RUN mvn clean package

2. This step results in an uber-jar (function.jar) that you can use as an input argument for jdeps. The output is a file containing all the Java modules that the function depends on:

RUN jdeps -q \
    --ignore-missing-deps \
    --multi-release 18 \
    --print-module-deps \
    target/function.jar > jre-deps.info

3. Create an optimized Java runtime based on those application modules with jlink. Remove unnecessary information from the runtime, for example header files or man-pages:

RUN jlink --verbose \
    --compress 2 \
    --strip-java-debug-attributes \
    --no-header-files \
    --no-man-pages \
    --output /jre18-slim \
    --add-modules $(cat jre-deps.info)

4. This creates your own custom Java 18 runtime in the /jre18-slim folder. You can apply additional optimization techniques such as Class-Data-Sharing (CDS) to generate a classes.jsa file to accelerate the class loading time of the JVM.

RUN /jre18-slim/bin/java -Xshare:dump

Adding optimized starting instructions

You must tell the Lambda execution environment how to start the application. You can achieve that with a bootstrap file that includes the necessary instructions. In addition, you can define parameters to improve the performance further. For example, you could use tiered compilation and SerialGC.

The following snippet represents an example of a bootstrap file:

#!/bin/sh

$LAMBDA_TASK_ROOT/jre18-slim/bin/java \
    --add-opens java.base/java.util=ALL-UNNAMED \
    -XX:+TieredCompilation \
    -XX:TieredStopAtLevel=1 \
    -XX:+UseSerialGC \
    -jar function.jar "$_HANDLER"

Packaging the components

Combine the bootstrap file, the custom Java runtime, and the application code in a zip file for later use as the deployment package:

RUN zip -r runtime.zip \
    bootstrap \
    function.jar \
    /jre18-slim

The GitHub example provides a build.sh script to run the above-mentioned process via Docker. This results in a runtime.zip that you can then use as a deployment package.

Deploying the application with the custom runtime

To deploy the custom runtime, use AWS CDK. This allows you to define the needed infrastructure as code more easily in your favorite programming language.

The following code snippet shows how to create a Lambda function from a custom runtime:

Function customJava18Function = new Function(this, "LambdaCustomRuntimeJava18", FunctionProps.builder()
        .functionName("custom-runtime-java-18")
.handler("com.amazon.aws.example.ExampleDynamoDbHandler::handleRequest")
        .runtime(Runtime.PROVIDED_AL2)
        .code(Code.fromAsset("../runtime.zip"))
        .memorySize(512)
        .environment(Map.of("TABLE_NAME", exampleTable.getTableName()))
        .timeout(Duration.seconds(20))
        .logRetention(RetentionDays.ONE_WEEK)
        .build());

To deploy the application and output the necessary API Gateway URL to invoke the Lambda function, use the following command or use the provided provision_infrastructure.sh script:

cdk deploy --outputs-file target/outputs.json

Testing the application and validating the example results

After deployment, you can load test the application with the open-source software project Artillery.

The following command creates 120 concurrent invocations of the Lambda function for a duration of 60 seconds. It uses the API Gateway URL that is exported after the AWS CDK successfully deployed the application:

artillery run -t $(cat infrastructure/target/outputs.json | jq -r '.LambdaCustomRuntimeMinimalJRE18InfrastructureStack.apiendpoint') -v '{ "url": "/custom-runtime" }' infrastructure/loadtest.yml

Use CloudWatch Log Insights to query the Lambda logs and gather information about the cold start (initDuration) and duration percentiles:

filter @type = "REPORT"
    | parse @log /\d+:\/aws\/lambda\/(?<function>.*)/
    | stats
    count(*) as invocations,
    pct(@duration+coalesce(@initDuration,0), 0) as p0,
    pct(@duration+coalesce(@initDuration,0), 25) as p25,
    pct(@duration+coalesce(@initDuration,0), 50) as p50,
    pct(@duration+coalesce(@initDuration,0), 75) as p75,
    pct(@duration+coalesce(@initDuration,0), 90) as p90,
    pct(@duration+coalesce(@initDuration,0), 95) as p95,
    pct(@duration+coalesce(@initDuration,0), 99) as p99,
    pct(@duration+coalesce(@initDuration,0), 100) as p100
    group by function, ispresent(@initDuration) as coldstart
    | sort by coldstart, function

The results provide an indication of how your application performs with the custom runtime. This is especially helpful when comparing different versions.

  • Invocation time (@duration) for both cold and warm starts plus function initialization time (@initDuration) if it is a cold start:

Invocation time

  • Function initialization time (@initDuration) only:

Function initialisation time

Conclusion

In this blog post, you learn how to create your own optimized Java runtime for AWS Lambda by using a variety of Java optimization techniques. This allows you to tailor your Java runtime to your application needs.

See the full example on GitHub and make use of your own preferred Java version. Add additional optimization steps in the Dockerfile or tune the parameters in the bootstrap file to optimize the start of the Java virtual machine.

In case you want to re-use your custom runtime in multiple Lambda functions, you can also distribute it via a Lambda layer.

For more serverless learning resources, visit Serverless Land.