How to enable Docker layer caching between builds in GitHub actions?

[sschuberth]

I was trying to speed up our Batect runs across PRs, where we're using Batect to setup an environment for functional tests. If nothing in the test environment / Dockerfile changes, my goal would be that Batect's setup across PRs would essentially be a no-op as everything comes from the layer cache.

However, guessing from the build time for funTest-analyzer-docker of a whopping 41 minutes, which is about twice as much as expected, Batect does not seem to make use of the layers restored via docker/build-push-action, and instead the Docker image was built twice.

So I'm wondering whether anyone was able to setup layer changing for Batect via https://github.com/docker/build-push-action/blob/master/docs/advanced/cache.md#github-cache?

[charleskorn]

Do you see the same behaviour (slower than expected image builds) when using docker build ... to build your image?

Also, if you're wanting to use the cache backend, keep in mind that you need to use cache-from and cache-to, which Batect doesn't yet support.

[charleskorn]

Another thing to try would be to enable more verbose output from Batect (--output all) and see if the output from the image build shows it rebuilding the image or reusing the existing layers.

[sschuberth]

if you run docker build . immediately after that step, does it reuse the restored layers?

It seems so, yes. See how here the "Build the Docker image" step takes 10 minutes whereas here in a subsequent run it takes only 8 seconds thanks to the reuse of cached layers.

Edit: I just realized that comparison to be incorrect, as it exclusively uses the docker/build-push-action which naturally can restore cache-from: type=gha. I'll do another test with running docker build . manually.

Edit 2: Indeed, this run shows that while a subsequent run of docker/build-push-action does get sped up thanks to caching, the manual docker build . run does not:

[sschuberth]

I've asked about guidance at docker/build-push-action#681.

[charleskorn]

OK, that's the behaviour I expected. Adding support for cache-from and cache-to isn't particularly difficult now - watch this space :)

[sschuberth]

For completeness, I've filed batect/docker-client#244 about this.

[colt-jay]

you need to use cache-from and cache-to, which Batect doesn't yet support.

@charleskorn I've reviewed the roadmap, issues, and PRs and it looks like this is not a priority right now right?

To make sure I understand, you suggest that we pre-build the images using docker build + the buildx caching mechanism so that batect can then build from that image?

[charleskorn]

Support for cache-from and cache-to will be added as soon as I have the new Docker client in place, which is my current focus area. I can't commit to a particular timeframe, but it will be soon :)

To make sure I understand, you suggest that we pre-build the images using docker build + the buildx caching mechanism so that batect can then build from that image?

That should work - Docker should pull down the layers it can reuse from the cache while it's building the image, and then Batect should just reuse the already-built image. If this isn't working for you, try running Batect with --output=all to get the output from the image build process to see if it's reusing layers from the local cache or not.

[sschuberth]

Support for cache-from and cache-to will be added as soon as I have the new Docker client in place

So, should that be the case with Batect 0.80.0 now?

[charleskorn]

Not yet :) But now that the new Docker client is in place, adding support for cache-from and cache-to should be straightforward.

[sschuberth]

I'm still wondering why there needs to be explicit support for cache-from / cache-to in Batect's Docker client. My understanding is still that the docker/build-push-action already restores Docker layers from the GitHub Action cache, and any tool that uses Docker afterwards should benefit from those cached layers...

Edit: I just realized that I never fully replied to your follow-up question above. Will do that now.

[sschuberth]

Meanwhile, I was also trying to pass load: true as suggested here to docker/build-push-action as the docs sounded as if that would make restored layers cached in gha available to CLI commands like docker build (and thus hopefully also to batect), but I couldn't get it to work. Images / layers seems to be rebuilt all the time.