balena-cli is not using the cache on fresh machine


i want to share builds with other people in my team. Therefore i set up a docker registry that stores the build images. Sadly when other people are pulling the image the cache seems not to be used.
Since our builds take ~3h on a fast workstation (CUDA, OpenCV, Tensorflow, …) caching is very usefull.

How to reproduce the behaviour

  1. push a build image to you registry
    docker tag myapp_main
    docker push
  2. prune all local images
    docker system prune --all
  3. pull image
    docker pull
    docker tag myapp_main:latest
  4. do a new build without changing the dockerfile or its resources
    -> Cache is not used

Do you have any idea on how to tag the images, in order to get chaching to work?

Can you provide us with more details on how you build the image in the first place and when the cache is used? Also is this an emulated build? In which case we start a docker daemon in a qemu to perform cross architecture builds.

I build the docker image cross platform. Therefore the qemu platform is used to build it.

Build and deploy in one stage:
balena deploy myapp --logs --dockerfile Dockerfile --source . --emulated --build

Is there a way to get the cached images into the ~/.balena/bin folder since it seems like the qemu builds are stored there or am i completely off?

Hi Frederic, the built images would be stored in your local docker registry once they are built. But I think most likely the CLI won’t by default use all the on disc images that exist while building. docker used to do this automatically but due to some cache poisoning exploits they disabled automatically using all images and only use images that were build by the engine on your host. So one needs to pass a specially flag that tells the build to use all the available images, or a list of images as cache. I suspect the cli deploy command is not using that flag when building, but it should be possible to add.

@shaunmulligan Can you point me in a direction on what i should look into in order to modify the balena-cli, to do so?

I think it would be a matter of adding it to . We would need to try add the --cache-from option to the build (more details on that option here: and it seems the images would need to be built using the buildkit backend of docker too, but unsure on that). I have asked the CLI maintainers to weigh in on this as well as they would be able to give more specific details on how to add that option.

The --cache-from option is what i looked into, too. But i just could not figure out, where i could hardcode it into the balena-cli for a quick fix and later do a Pull Request with a better implementation. I am no expert in coffee script and the project seems quiet complex to understand. Therefore i am looking for file you can point me to, that describes how to sett the build options with docerode or similar, so i can work my way from there.

Yeah, I have been trying to find where the options are set as well, but have had no luck and also have little to zero coffee experience, so think we need to wait until the CLI team come online, they will be able to point us in the right direction super fast.

Cool, thank you for your help. Looking forward to hear from the CLI team :slight_smile:

Just as information for the CLI team. The build argument needs to be parsed through dockerode. Since it isusing the docker api docker_build
The cachefrom option is defined as string: JSON array of images used for build cache resolution..

Therefore i am looking for file you can point me to, that describes how to set the build options with dockerode or similar, so i can work my way from there.

@Langhalsdino, I’ve created a draft PR, not yet tested and thus maybe not working, with the likely changes to add a --cachefrom option to the balena build and deploy commands:

If you get to test it, your feedback will be very welcome!

1 Like

I will try it out in the next hours, since this feature is very important to us :slight_smile:
We spend about 8h a day on building images and this feature can cut it down to ~1h a day :slight_smile:

Yes i tried it and it does not seem to work. My current guess is, that the Dockerfile will get treated similar to how docker compose gets treated, therefore we might need to implement something like this.

More of the details can be found here

I do not know why, but it started working now with docker-compose files and regular dockerfiles with no change at all :smiley:

Thanks for the PR :slight_smile:

Just as a recap for other who might be interested:

  1. Clone the repository and build the app
git clone
cd balena-cli
git checkout 1616-deploy-cachefrom
npm install
npm rund build:fast
export MY_PATH_TO_BALENA=$(pwd)
  1. build your app with caching
cd your/balena/project/path
docker pull
${MY_PATH_TO_BALENA}/balena-cli/bin/balena deploy my-app --logs --dockerfile Dockerfile --source . --emulated --build-arg BUILDKIT_INLINE_CACHE=1 --cache-from -t
  1. push new build to registry
docker push

Thanks for posting your results, I will inform the maintainer that the modification works for you.

@Langhalsdino, thank you for the feature suggestion and for helping with the testing. The --cache-from option for the balena build and balena deploy commands was released in CLI version 11.26.0.
For the benefit of others reading this thread, it implements the same feature documented for the docker build --cache-from option, “Specifying external cache sources”: