Azure DevOps pipeline with multi-stage docker image layer caching support

azure-pipelines-build-docker-with-cache.yml

trigger:
- master

pool:
  vmImage: 'ubuntu-latest'

variables:
  imageRepository: 'multi-stage-build-caching'
  containerRegistry: 'ACR-Service-Connection-Name'
  registryName: '[redacted].azurecr.io'
  tag: '$(Build.BuildNumber)'

steps:
  - task: Docker@2
    displayName: "Registry Login"
    inputs:
      containerRegistry: '$(containerRegistry)'
      command: 'login'

  - script: "docker pull $(registryName)/$(imageRepository):latest-build && docker pull $(registryName)/$(imageRepository):latest"
    displayName: "Pull latest for layer caching"
    condition: ne(variables['USE_BUILD_CACHE'], 'false')
    continueOnError: true # for the first build, tags will not exist yet

  - task: Docker@2
    displayName: "Build - Builder Stage"
    inputs:
      containerRegistry: '$(containerRegistry)'
      repository: '$(imageRepository)'
      command: 'build'
      Dockerfile: './Dockerfile'
      buildContext: '.'
      arguments: '--target builder --cache-from=$(registryName)/$(imageRepository):latest-build'
      tags: |
        $(tag)-build
        latest-build

  - task: Docker@2
    displayName: "Build - Final Stage"
    inputs:
      containerRegistry: '$(containerRegistry)'
      repository: '$(imageRepository)'
      command: 'build'
      Dockerfile: './Dockerfile'
      buildContext: '.'
      arguments: '--cache-from=$(registryName)/$(imageRepository):latest-build --cache-from=$(registryName)/$(imageRepository):latest'
      tags: |
        $(tag)
        latest

  - task: Docker@2
    displayName: "Push image"
    inputs:
      command: push
      containerRegistry: "$(containerRegistry)"
      repository: $(imageRepository)
      tags: |
        $(tag)-build
        latest-build
        $(tag)
        latest

@girip11 this solution definitely did work for me, and I believe it's still in use at the company I did it for. The problem is that the issue you're running into is different from what I was solving here.

Your sample (similar to the SO thread) is expecting to skip building the test target because it's not referenced by the production final target. As the only answer on SO points out - docker doesn't support skipping stages, but if you use buildkit it looks to be possible (I haven't tried this myself).

My sample here is a workaround to MS Build Agents being ephemeral - so there are no layers cached each time a new build runs. By tagging, and pushing/pulling the latest-build tag I was able to avoid time spent rebuilding all the layers every time the build ran because docker would see the layer as not having any changes vs what was already built. Properly implemented you'll see messages in your build log on those steps saying "using cache" and the work won't be redone.

@TheDruidsKeeper Thanks for the reply. You are right. In my case I had to build test and production separately(ideally I wanted the test to use production as base). Using the buildkit works. I just have to set the environment variable DOCKER_BUILDKIT=1. Docker recent versions support skipping stages, but we were using custom agent pools where the docker version was old. Yes I did enable docker caching and it works. Once again thanks a lot for responding.

Great to hear 👍