DEPLOYING GO APPS ON DOCKER SCRATCH IMAGES

https://weberc2.bitbucket.io/posts/golang-docker-scratch-app.html

2018-11-04

A few months ago I built out some monitoring infrastructure at work using Go. I deployed it to ECS (a Docker orchestrator, functionally similar to Kubernetes), and for fun I decided to see how minimal I could make the image. I’ve used Alpine base images before (which weigh in at about 5 MB and usually another 5 MB for a small Go binary), but being that Go advertises itself as requiring only a Linux kernel (most programming languages depend on an interpreter, a VM, and/or system libraries–the latter abstract over kernel functionality and sometimes provide a stable interface to the kernel), I wanted to see how true or practical this was, and I wanted to better understand the things that I was taking for granted by using distros.

As a matter of context, Docker has a special base image called scratchwhich is empty–an application running on a scratch base image only has access to the kernel (at least to the extent that containers provide isolation).

NOTE: This approach will not work if your Go application needs to subprocess out to other programs (e.g., git) or if it uses pretty much any library that uses CGo (these almost always depend on libc if not other libaries).

NOTE: The build steps may vary if you’re not using Go modules.

Typically, to build a Docker image for a Go application, you write a Dockerfile that specifies a base Docker image with the Go toolchain installed, copies in the source code from the host machine, and invokes the compiler against it to produce the binary before committing the changes to the final image. It looks something like this:

FROM golangWORKDIR /workspace# Assuming the source code is collocated to this Dockerfile, copy the whole# directory into the container that is building the Docker image.COPY . .RUN go build -o /myapp# When a container is run from this image, run the binaryCMD /myapp

The toolchain and its dependencies (git, mercurial, etc) weigh a few hundred MBs (never mind the weight of the distribution itself), so it’s common to use a Docker feature called multi-stage Builds to copy the binary artifact from this first image into a second image without the toolchain. So we now have something like this:

FROM golangWORKDIR /workspaceCOPY . .RUN go build -o /myappFROM scratch# Copy the artifact from the first build stage into the second stage (which will# become the final image)COPY --from=0 /myapp /myapp

Note that we’ve removed the CMD /myapp line because (probably for good reasons that I don’t fully understand), that actually runs /bin/sh -c /myapp, so running the Docker image with the default command would otherwise give the error: docker: Error response from daemon: OCI runtime create failed: container_linux.go:348: starting container process caused "exec: \"/bin/sh\": stat /bin/sh: no such file or directory": unknown (which is particularly unhelpful). At least now, running the image without specifying a command will give a more helpful error: docker: Error response from daemon: No command specified.

This will work for something like hello world, but real apps have a few more requirements. First of all, some packages in the standard library will try to link against system libraries by default (IIRC, net specifically prefers to use the system DNS resolver where possible). This means that containers created from this image will fail at runtime because the binary can’t find the system libraries (because they don’t exist). However, pure-Go implementations will be used if the compiler is invoked with CGO_ENABLED=0:

FROM golangWORKDIR /workspaceCOPY . .RUN CGO_ENABLED=0 go build -o /myappFROM scratchCOPY --from=0 /myapp /myapp

Now we can get past the dynamic linking error, but we run into another curiosity if our application needs to make HTTPS requests (or pretty much anything else that needs to use SSL): Get https://www.google.com: x509: certificate signed by unknown authority. This error comes from net/http–not from Docker. We can avoid this error by not handling it in the program: rsp, _ := http.Get(url). Kidding. This error is telling us that the HTTP library can’t find the certs required to establish an SSL connection (or something–I really understand SSL much less than I probably should). Basically scratch also doesn’t come with these certs, so how do we get them into our final image? Copy them from the builder stage!

FROM golangWORKDIR /workspaceCOPY . .RUN CGO_ENABLED=0 go build -o /myappFROM scratchCOPY --from=0 /myapp /myappCOPY --from=0 /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/

Now the application should run fine; however, by default, Docker images are run as root user–to follow security best practices, we should use a nonroot user. However, scratch doesn’t have programs like adduser nor even echo, so we can’t echo “\$USER_INFO” >> /etc/passwd, but we can create a file in the builder stage and copy that over to the final stage:

# This is the first stage, for building things that will be required by the# final stage (notably the binary)FROM golang# Assuming the source code is collocated to this DockerfileCOPY . .# Build the Go app with CGO_ENABLED=0 so we use the pure-Go implementations for# things like DNS resolution (so we don't build a binary that depends on system# libraries)RUN CGO_ENABLED=0 go build -o /myapp# Create a "nobody" non-root user for the next image by crafting an /etc/passwd# file that the next image can copy in. This is necessary since the next image# is based on scratch, which doesn't have adduser, cat, echo, or even sh.RUN echo "nobody:x:65534:65534:Nobody:/:" > /etc_passwd# The second and final stageFROM scratch# Copy the binary from the builder stageCOPY --from=0 /myapp /myapp# Copy the certs from the builder stageCOPY --from=0 /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/# Copy the /etc_passwd file we created in the builder stage into /etc/passwd in# the target stage. This creates a new non-root user as a security best# practice.COPY --from=0 /etc_passwd /etc/passwd# Run as the new non-root by defaultUSER nobody

It turns out this is all we need to do. The certs and /etc/passwd file are negligible in terms of size, so the final image is dominated by the size of your executable. My small executable weighed 4.5 MB uncompressed and ECR (Amazon’s Docker Hub analog) reported it as 2.5 MB (presumably this is compressed). There are a number of tricks for reducing your binary size further, but I’ll leave those as an exercise for the reader.

There is one last issue with this Dockerfile: rebuilds will have to pull the dependencies every time, which takes a while (especially tedious if you’re iterating on the Dockerfile itself). To ameliorate this, we’ll add just the go.mod and go.sum dependencies and run go mod download before we copy in the rest of the source code–in so doing, the Docker build cache will only redownload dependencies when either the go.mod or go.sum files have changed:

# This is the first stage, for building things that will be required by the# final stage (notably the binary)FROM golang# Copy in just the go.mod and go.sum files, and download the dependencies. By# doing this before copying in the other dependencies, the Docker build cache# can skip these steps so long as neither of these two files change.COPY go.mod go.sum ./RUN go mod download# Assuming the source code is collocated to this DockerfileCOPY . .# Build the Go app with CGO_ENABLED=0 so we use the pure-Go implementations for# things like DNS resolution (so we don't build a binary that depends on system# libraries)RUN CGO_ENABLED=0 go build -o /myapp# ...

This wasn’t much more work than building off of an Alpine base image, and I deployed it this way. The size advantage over Alpine is still only ~5 MB, and while that’s a significant percentage, the impact on your workflow or deployments will hardly be perceptible (largely due to caching). The biggest reason to choose scratch over alpine is security–reduced attack surface and all that–which may be a more practical concern than I originally thought.

CONTACT

# This is the first stage, for building things that will be required by the# final stage (notably the binary)FROM golang# Copy in just the go.mod and go.sum files, and download the dependencies. By# doing this before copying in the other dependencies, the Docker build cache# can skip these steps so long as neither of these two files change.COPY go.mod go.sum ./RUN go mod download# Assuming the source code is collocated to this DockerfileCOPY . .# Build the Go app with CGO_ENABLED=0 so we use the pure-Go implementations for# things like DNS resolution (so we don't build a binary that depends on system# libraries)RUN CGO_ENABLED=0 go build -o /myapp# Create a "nobody" non-root user for the next image by crafting an /etc/passwd# file that the next image can copy in. This is necessary since the next image# is based on scratch, which doesn't have adduser, cat, echo, or even sh.RUN echo "nobody:x:65534:65534:Nobody:/:" > /etc_passwd# The second and final stageFROM scratch# Copy the binary from the builder stageCOPY --from=0 /myapp /myapp# Copy the certs from the builder stageCOPY --from=0 /etc/ssl/certs/ca-certificates.crt /etc/ssl/certs/# Copy the /etc_passwd file we created in the builder stage into /etc/passwd in# the target stage. This creates a new non-root user as a security best# practice.COPY --from=0 /etc_passwd /etc/passwd# Run as the new non-root by defaultUSER nobody