Optimising cargo build times in Docker

Question

I am developing an API with Rust, and am managing the environments, including the external database with Docker. Every time I make a change to the API code, cargo rebuilds, and since Docker doesn't cache anything to do with the ADD statement to copy the Rust directory over to the container, it re-downloads all the packages, which is a fairly lengthy process since I'm using Nickel, which seems to have a boatload of dependencies.

Is there a way to bring those dependencies in prior to running cargo build? At least that way if the dependencies change it will only install what's required, similar to Cargo compiling locally.

Here's the Dockerfile I currently use:

FROM ubuntu:xenial
RUN apt-get update && apt-get install curl build-essential ca-certificates file xutils-dev nmap -y
RUN mkdir /rust
WORKDIR /rust
RUN curl https://sh.rustup.rs -s >> rustup.sh
RUN chmod 755 /rust/rustup.sh
RUN ./rustup.sh -y
ENV PATH=/root/.cargo/bin:$PATH SSL_VERSION=1.0.2h
RUN rustup default 1.11.0
RUN curl https://www.openssl.org/source/openssl-$SSL_VERSION.tar.gz -O && \
    tar -xzf openssl-$SSL_VERSION.tar.gz && \
    cd openssl-$SSL_VERSION && ./config && make depend && make install && \
    cd .. && rm -rf openssl-$SSL_VERSION*
ENV OPENSSL_LIB_DIR=/usr/local/ssl/lib \
    OPENSSL_INCLUDE_DIR=/usr/local/ssl/include \
    OPENSSL_STATIC=1
RUN mkdir /app
WORKDIR /app
ADD . /app/
RUN cargo build
EXPOSE 20000
CMD ./target/debug/api

And here's my Cargo.toml

[profile.dev]
debug = true

[package]
name = "api"
version = "0.0.1"
authors = ["Vignesh Sankaran <[e-mail]>"]

[dependencies]
nickel = "= 0.8.1"
mongodb = "= 0.1.6"
bson = "= 0.3.0"
uuid = { version = "= 0.3.1", features = ["v4"] }

Answer 1

You can create an intermediate image and build your final image from it. Eg:

FROM ubuntu:xenial
RUN apt-get update && apt-get install curl build-essential ca-certificates file xutils-dev nmap -y
RUN mkdir /rust
...

build using docker build -t mybaseimage .

FROM mybaseimage
RUN mkdir /app
WORKDIR /app
ADD . /app/
RUN cargo build
EXPOSE 20000
CMD ./target/debug/api

docker build -t finalimage .

That way only the mybaseimage is rebuilt

Answer 2

Docker does cache the layer built from the ADD (preferably COPY) instruction, provided the sources haven't changed. You could make use of that and get your dependencies cached by copying the Cargo.toml in first, and doing a build.

But unfortunately you need something to build, so you could do it with a single source file and a dummy lib target in your manifest:

[lib]
name = "dummy"
path = "dummy.rs"

In your Dockerfile build the dummy separately:

COPY Cargo.toml /app/Cargo.toml
COPY dummy.rs /app/dummy.rs
RUN cargo build --lib

The output of this layer will be cached, with all the dependencies installed, and then you can go on to add the rest of your code (in the same Dockerfile):

COPY /src/ app/src/
RUN cargo build

The dummy stuff is ugly, but it means your normal build will be quick, as it comes from the cached layer, and when you change dependencies in your Cargo.toml then Docker will pick it up and build a new layer with updated dependencies.

Answer 3

This question is a year and a half old at this point. Still, no cargo build --deps-only option, but I thought I'd share my solution, which is reasonably lightweight. You don't have to modify any of your host files to do it:

COPY Cargo.toml .
RUN mkdir src \
    && echo "// dummy file" > src/lib.rs \
    && cargo build

This will build the dependencies and cache them. Later when you copy in the actual source files (or in my case, use --volumes), it will overwrite the dummy file, so the dummy file is totally temporary. You could also explicitly rm it after the build if you needed to.

Answer 4

Instead of adding a dummy file, you can also let the build fail:

RUN cargo build || true
COPY ...
RUN cargo build

Don't forget to add --release to both places if you want optimized builds.