Problem

I faced a peculiar problem where a fat-jar wouldn't run on remote runners. I described the details in my previous post - Dataflow + SpringBoot app fails to run when Dockerized.

To unblock myself and deliver on time, I hacked together a solution that during the build step instead of compiling the source to create a jar and containerize the jar, I packaged the source repository and used the gradle wrapper to run that code when the container starts.
It worked but it had a few problems:

1. ❗️The images were almost 1 GB. Our typical service is around 350 MB out of which 195 MB is the Distroless Java 11 itself.
2. ❗️The startup was slow as gradle was starting first and then booting the app.
3. ❗️The image wasn't self-sufficient and I had to inject Artifactory credentials at runtime to start the container so that gradle can authenticate and validate dependencies.

All of these made the solution less than ideal.

In the following days, when I had more time, I kept looking into the issue and found the following issues on Github and Apache BEAM project

• Files from classpath are not properly resolved when classpath JAR contains META-INF with references to other dependencies
• DataflowRunner support for Class-Path jars

The above seemed related but still focused on different problems and not related specifically to SpringBoot so I also created a 🐞 BUG ticket with details.

Solution

Reading through different comments, I saw a suggestion to use maven-shade-plugin. We use gradle and I learnt that there is something similar for it called Shadow Gradle plugin

Sharing what it is directly from their page:

Shadow is a Gradle plugin for combining a project's dependency classes and resources into a single output Jar. The combined Jar is often referred to a fat-jar or uber-jar. Shadow utilizes JarInputStream and JarOutputStream to efficiently process dependent libraries into the output jar without incurring the I/O overhead of expanding the jars to disk.

Shadowing a project output has 2 major use cases:

1. Creating an executable JAR distribution.
2. Bundling and relocating common dependencies in libraries to avoid classpath conflicts

👨🏻‍💻I gave it a try and it worked. ✨

Configuration

This is my build.gradle.kts configuration:

// shadow config
tasks.withType<ShadowJar> {
    isZip64 = true
    // Required for Spring
    mergeServiceFiles()
    append("META-INF/spring.handlers")
    append("META-INF/spring.schemas")
    append("META-INF/spring.tooling")
    transform(PropertiesFileTransformer().apply {
        paths = listOf("META-INF/spring.factories")
        mergeStrategy = "append"
    })
}

With this configuration in place, the build pipeline runs ./gradlew build in the step to compile and create the merged jar. Shadow compiles the file to a jar with -all.jar suffix which is then copied to the Docker image.
This is the Dockerfile:

FROM gcr.io/distroless/java:11
VOLUME /tmp
COPY build/libs/dataExtractor-0.0.1-all.jar app.jar
EXPOSE 8080
ENTRYPOINT ["java", "-Djava.security.egd=file:/dev/./urandom", "-Dspring.profiles.active=${ENV_SPRING_PROFILE}", "-jar", "/app.jar"]

This solved the 3 problems:

• ✅ The image is 340 MB and that is close to expected.
• ✅ Startup times are similar to expected.
• ✅ No more injected Artifactory credentials. I rotated the credentials that were used for the hack.

This was an interesting problem because the app worked every time from the IDE or when run locally via gradle but failed when it was containerized and deployed.

Please let me know if you think there is something that can be improved.

Originally published on https://suspendfun.com