[Build] Reduce pulsar-test-latest-version docker image size

[lhotari]

Motivation

Smaller pulsar-test-latest-version docker image size makes it feasible to build the image once and share it between the different parallel integration test build jobs. This will help improve the Pulsar CI in the future since it opens up new possibilities to save build system resources (GitHub Actions Runner VM time).

Modifications

• Build go-lang examples in a separate temporary layer. This reduces the final size of the image by at least 600MB.

• Cleanup files as last steps

  • Cleaning up files doesn't reduce the size of the image yet since it requires docker build squashing. The cleanup step is preparation for future changes.
  • To actually reduce the final image size, it would require -Ddockerfile.build.squash=true (uses docker build --squash). This is not part of this change since the image size isn't important until images are transferred over the network. Squashing the image would just add overhead to the current build.
  • Remove Pulsar CPP library's static libraries /usr/lib/libpulsar*.a (since no tests use it)
  • Remove gcc compiler (since no tests use it)

[eolivelli]

+1

[zymap]

I wonder how to share the docker image between different jobs?

[lhotari]

I wonder how to share the docker image between different jobs?

@zymap The high level idea is described in the draft PIP Changes to GitHub Actions based Pulsar CI. It's possible by using the GitHub Actions Cache. There's an action that adds a local HTTP proxy to the cache where you can POST and GET any key from the cache. I've been experimenting with it and it's really great and efficient. I'll share the results on the mailing list asap (next week).

docker comes with docker save and docker load commands which allow to export an image to a stream of bytes. I've been experimenting with a solution where I pipe this to zstd compressor and with curl over http-cache-proxy to GitHub Actions Cache. It's pretty stable, however sometimes the upload or download fails. I'll add retries to address that issue.

btw. The challenge with docker save is that it saves all layers. It also saves the openjdk:8-jdk-slim layers which could be pulled directly, however that's not a big concern since the size of openjdk:8-jdk-slim is 288MB uncompressed. This is a small portion of the total image size of pulsar-test-latest-version image. There are solutions to hack around the limitation, but they are fairly complex.

[lhotari]

/pulsarbot run-failure-checks

[lhotari]

/pulsarbot run-failure-checks

[michaeljmarshall]

@lhotari - great change, and I'm looking forward to seeing the GitHub Actions Cache in practice. That should really speed up the builds.