chore: move go mod download before copy source in Dockerfile

[fscnick]

Why are these changes needed?

To avoid downloading dependences after changing the source file, move go mod download up before copying source files.

In docker build, the cache depends on the previous layer. Which means the upper layer has changed, the following layers are subject to re-build.

Usually, the modification of deps is less frequent than the modification of source code. In the origin Dockerfile,

kuberay/apiserver/Dockerfile

Lines 9 to 17 in 37cf2ac

	# Copy the go source
	COPY proto/ proto/
	COPY ray-operator/ ray-operator/
	COPY apiserver/ apiserver/
	WORKDIR /workspace/apiserver
	# cache deps before building and copying source so that we don't need to re-download as much
	# and so that source changes don't invalidate our downloaded layer
	RUN go mod download

go mod download is after copying the go source. Once source files has changed. the docker build will re-run go mod download without leveraging the cache.

Thus, move go mod download before the copying the go source.

Related issue number

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

[dentiny]

To avoid downloading dependences after changing the source file

Sorry I'm not that familiar with golang, just for my own curiosity, could you please elaborate a little bit on the problem you've seen? thank you!

[fscnick]

To avoid downloading dependences after changing the source file

Sorry I'm not that familiar with golang, just for my own curiosity, could you please elaborate a little bit on the problem you've seen? thank you!

• go.mod and go.sum are the deps file in Go. And go mod download fetch the deps.
• In docker build, the cache depends on the previous layer. Which means the upper layer has changed, the following layers are subject to re-build.

Usually, the modification of deps is less frequent than the modification of source code. In the origin Dockerfile,

kuberay/apiserver/Dockerfile

Lines 9 to 17 in 37cf2ac

	# Copy the go source
	COPY proto/ proto/
	COPY ray-operator/ ray-operator/
	COPY apiserver/ apiserver/
	WORKDIR /workspace/apiserver
	# cache deps before building and copying source so that we don't need to re-download as much
	# and so that source changes don't invalidate our downloaded layer
	RUN go mod download

go mod download is after copying the go source. Once source files has changed. the docker build will re-run go mod download without leveraging the cache.

Thus, move go mod download before the copying the go source. In ray-operator, it is right after copy deps.

kuberay/ray-operator/Dockerfile

Lines 6 to 10 in 37cf2ac

	COPY go.mod go.mod
	COPY go.sum go.sum
	# cache deps before building and copying source so that we don't need to re-download as much
	# and so that source changes don't invalidate our downloaded layer
	RUN go mod download

In addition, it copies ray-operator's and proto's go.mod and go.sum in the dockerfile because apiserver/go.mod replaces the module with local directory.

kuberay/apiserver/go.mod

Lines 96 to 99 in 37cf2ac

	replace (
	github.com/ray-project/kuberay/proto => ../proto
	github.com/ray-project/kuberay/ray-operator => ../ray-operator
	)

Kindly let me know if there is something unclear or wrong.

[dentiny]

Thanks for the detailed explanation! I think it makes sense a lot.
Could you please at least add the docker layer part into PR description? I didn't get it at first sight. Thank you a ton!

[dentiny]

I feel like it's a recommended practice, to download all dependencies right after we figure out the go.mod`.

[kevin85421]

The change makes sense to me, but I'm curious about the difference between (1) the build time for updating the KubeRay API server source code without this PR, and (2) the build time for updating the KubeRay API server source code with this PR.

[fscnick]

The change makes sense to me, but I'm curious about the difference between (1) the build time for updating the KubeRay API server source code without this PR, and (2) the build time for updating the KubeRay API server source code with this PR.

Without this PR: (95.9s)

With this PR: (85.4s)

[kevin85421]

Thank you!