[Q] Jib CLI vs Jib Gradle plugin performance

[artemptushkin]

I want to get public opinion and know how others see and organize this.

Given:

• a big Kotlin/Java project - 5 minutes for ./gradlew assemble from scratch, many submodules
• CI/CD: build - test - build-image
• We consider here build-image job where we run build Docker image with jibDockerBuild and push it

The reality is that Gradle defines the chain tasks so we always pass redundant configuration steps, and we can easily get into a situation where we recompile something just because the task is not incremental from Gradle's perspective (another story).

What's this for pipeline?

We have to pull and push the Gradle cache for other non-significant tasks (compile) -> for the big project it's time-consuming

Assumption

If we get rid of the plugin and replace it with just Jib CLI jib jar the job will be faster keeping the same functionality

Practice

It's faster confirmed:

• We don't have redundant Java processes, Gradle tasks - so we can delete some cache pull/push

From 3-5 minutes of running the Gitlab CI/CD job we have now 1-2 minutes

But....

The startup changes from ~25 seconds to ~40 seconds, you can see the Spring log startup degradation on the screenshot

The layers/start difference:

• main-class:
  • Jib Gradle plugin: the Spring Boot main application class
  • Jib CLI: Spring Boot JarLaucher class
• docker entrypoint:
  • Jib Gradle plugin: [ "java", "-cp", "@/app/jib-classpath-file", "com.iptiq.apptemplate.AppTemplateApplicationKt" ]
  • Jib CLI: [ "java", "-cp", "@/app", "org.springframework.boot.loader.JarLauncher" ]
• layers:
  • Jib Gradle produces 5 layers:
  • third-party jars
  • application jars
  • classpath resources
  • classes
  • run files
  • Jib CLI produces 3 layers:
  • third-party jars
  • Spring Boot loader classes
  • application jars, classpath resources, classes, run files

Questions.

Why is Spring start-up time so different between Jib CLI and Jib Gradle plugin build?

Are there any known best practices to optimize the build and startup time?

Is Jib approach documented? I want to know better how and why you build like this

[wwadge]

Issue feels a bit all over the place. In my builds, gradle skips over all tasks that it knows it doesn't need to do work on, this includes jib plugin (pass the skip option) to avoid pushing images for things that haven't changed.

Re startup time and differences, I'm going to guess this is Spring boot's plugin adding additional stuff when done via a gradle task (which doesn't execute in the CLI). Things you may want to explore to improve your startup cost:

• AppCDS : around 10-15% speedup in my builds
• CRaC: 50ms startup time for a typical spring boot app (see: https://github.com/sdeleuze/spring-boot-crac-demo)

[artemptushkin]

I want to get back to the topic and say that for one of my projects I have set up jib cli (jar) without the Gradle plugin and it's amount ~2 times more efficient because we ignore all the gradle scanning and configurations as they are irrelevant.

This plugin only needs the JAR, which can be produced as an artifact of a separate job.

[wwadge]

Gradle can also be told to cache configurations too in gradle.properties if you want to see the same effect. Takes some effort to get everything cached correctly however.