Enabling Docker Layer Caching
This document offers an overview of Docker Layer Caching (DLC), which can reduce Docker image build times on CircleCI. Docker Layer Caching is only available on the Performance usage plan. DLC is available on the Premium usage plan, on all server-installations.
Note: Docker Layer caching is only available on select plans:
Overview
Docker Layer Caching (DLC) is a great feature to use if building Docker images is a regular part of your CI/CD process. DLC will save image layers created within your jobs, rather than impact the actual container used to run your job.
DLC caches the individual layers of any Docker images built during your CircleCI jobs, and then reuses unchanged image layers on subsequent CircleCI runs, rather than rebuilding the entire image every time. In short, the less your Dockerfiles change from commit to commit, the faster your image-building steps will run.
Docker Layer Caching can be used with both the machine executor and the Remote Docker Environment (setup_remote_docker).
Limitations
Note: DLC has no effect on Docker images used as build containers. That is, containers that are used to run your jobs are specified with the image key when using the docker executor and appear in the Spin up Environment step on your Jobs pages.
DLC is only useful when creating your own Docker image with docker build, docker compose, or similar docker commands), it does not decrease the wall clock time that all builds take to spin up the initial environment.
version: 2
jobs:
build:
docker:
# DLC does nothing here, its caching depends on commonality of the image layers.
- image: circleci/node:9.8.0-stretch-browsers
steps:
- checkout
- setup_remote_docker:
docker_layer_caching: true
# DLC will explicitly cache layers here and try to avoid rebuilding.
- run: docker build .
How DLC Works
DLC caches your Docker image layers by creating an external volume and attaching it to the instances that execute the machine and Remote Docker jobs. The volume is attached in a way that makes Docker save the image layers on the attached volume. When the job finishes, the volume is disconnected and re-used in a future job. This means that the layers downloaded in a previous job with DLC will be available in the next job that uses the same DLC volume.
One DLC volume can only be attached to one machine or Remote Docker job at a time. If one DLC volume exists but two jobs that request DLC are launched, CircleCI will create a new DLC volume and attach it to the second job. From that point on the project will have two DLC volumes associated with it. This applies to parallel jobs as well: if two machine jobs are run in parallel, they will get different DLC volumes.
Depending on which jobs the volumes are used in, they might end up with different layers saved on them. The volumes that are used less frequently might have older layers saved on them.
The DLC volumes are deleted after 14 days of not being used in a job.
CircleCI will create a maximum of 50 DLC volumes per project, so a maximum of 50 concurrent machine or Remote Docker jobs per project can have access to DLC. This takes into account the parallelism of the jobs, so a maximum of 1 job with 50x parallelism will have access to DLC per project, or 2 jobs with 25x parallelism, and so on.
Remote Docker Environment
To use DLC in the Remote Docker Environment, add docker_layer_caching: true under the setup_remote_docker key in your config.yml file:
- setup_remote_docker:
docker_layer_caching: true # default - false
Every layer built in a previous job will be accessible in the Remote Docker Environment. However, in some cases your job may run in a clean environment, even if the configuration specifies docker_layer_caching: true.
If you run many parallel jobs for the same project that depend on the same environment, all of them will be provided with a Remote Docker environment. Docker Layer Caching guarantees that jobs will have exclusive Remote Docker Environments that other jobs cannot access. However, some of the jobs may have cached layers, some may not have cached layers, and not all of the jobs will have identical caches.
Note: Previously DLC was enabled via the reusable: true key. The reusable key is deprecated in favor of the docker_layer_caching key. In addition, the exclusive: true option is deprecated and all Remote Docker VMs are now treated as exclusive. This means that when using DLC, jobs are guaranteed to have an exclusive Remote Docker Environment that other jobs cannot access.
Machine Executor
Docker Layer Caching can also reduce job runtimes when building Docker images using the machine executor. Use DLC with the machine executor by adding docker_layer_caching: true below your machine key (as seen above in our example):
machine:
docker_layer_caching: true # default - false
Examples
Let’s use the following Dockerfile to illustrate how Docker Layer Caching works. This example Dockerfile is adapted from our Elixir convenience image:
Dockerfile
FROM elixir:1.6.5
# make apt non-interactive
RUN echo 'APT::Get::Assume-Yes "true";' > /etc/apt/apt.conf.d/90circleci \
&& echo 'DPkg::Options "--force-confnew";' >> /etc/apt/apt.conf.d/90circleci
ENV DEBIAN_FRONTEND=noninteractive
# man directory is missing in some base images
# https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=863199
RUN apt-get update \
&& mkdir -p /usr/share/man/man1 \
&& apt-get install -y \
git mercurial xvfb \
locales sudo openssh-client ca-certificates tar gzip parallel \
net-tools netcat unzip zip bzip2 gnupg curl wget
# set timezone to UTC
RUN ln -sf /usr/share/zoneinfo/Etc/UTC /etc/localtime
# use unicode
RUN locale-gen C.UTF-8 || true
ENV LANG=C.UTF-8
# install docker
RUN set -ex \
&& export DOCKER_VERSION=$(curl --silent --fail --retry 3 \
https://download.docker.com/linux/static/stable/x86_64/ | \
grep -o -e 'docker-[.0-9]*-ce\.tgz' | sort -r | head -n 1) \
&& DOCKER_URL="https://download.docker.com/linux/static/stable/x86_64/${DOCKER_VERSION}" \
&& echo Docker URL: $DOCKER_URL \
&& curl --silent --show-error --location --fail --retry 3 --output /tmp/docker.tgz "${DOCKER_URL}" \
&& ls -lha /tmp/docker.tgz \
&& tar -xz -C /tmp -f /tmp/docker.tgz \
&& mv /tmp/docker/* /usr/bin \
&& rm -rf /tmp/docker /tmp/docker.tgz
# install docker-compose
RUN curl --silent --show-error --location --fail --retry 3 --output /usr/bin/docker-compose \
https://circle-downloads.s3.amazonaws.com/circleci-images/cache/linux-amd64/docker-compose-latest \
&& chmod +x /usr/bin/docker-compose \
&& docker-compose version
# setup circleci user
RUN groupadd --gid 3434 circleci \
&& useradd --uid 3434 --gid circleci --shell /bin/bash --create-home circleci \
&& echo 'circleci ALL=NOPASSWD: ALL' >> /etc/sudoers.d/50-circleci \
&& echo 'Defaults env_keep += "DEBIAN_FRONTEND"' >> /etc/sudoers.d/env_keep
USER circleci
CMD ["/bin/sh"]
Config.yml
In the config.yml snippet below, let’s assume the build_elixir job is regularly building an image using the above Dockerfile. By adding docker_layer_caching: true underneath our machine executor key, we ensure that CircleCI will save each Docker image layer as this Elixir image is built.
version: 2
jobs:
build_elixir:
machine:
docker_layer_caching: true
steps:
- checkout
- run:
name: build Elixir image
command: docker build -t circleci/elixir:example .
On subsequent commits, if our example Dockerfile has not changed, then DLC will pull each Docker image layer from cache during the build Elixir image step, and our image will theoretically build almost instantaneously.
Now, let’s say we add the following step to our Dockerfile, in between the # use unicode and # install docker steps:
# install jq
RUN JQ_URL="https://circle-downloads.s3.amazonaws.com/circleci-images/cache/linux-amd64/jq-latest" \
&& curl --silent --show-error --location --fail --retry 3 --output /usr/bin/jq $JQ_URL \
&& chmod +x /usr/bin/jq \
&& jq --version
On the next commit, DLC will ensure that we still get cached image layers for the first few steps in our Dockerfile—pulling from elixir:1.6.5 as our base image, the # make apt non-interactive step, the step starting with RUN apt-get update, the # set timezone to UTC step, and the # use unicode step.
However, because our #install jq step is new, it and all subsequent steps will need to be run from scratch, because the Dockerfile changes will invalidate the rest of the image layer cache. Overall, though, with DLC enabled, our image will still build more quickly, due to the unchanged layers/steps towards the beginning of the Dockerfile.
If we were to change the first step in our example Dockerfile—perhaps we want to pull from a different Elixir base image—then our entire cache for this image would be invalidated, even if every other part of our Dockerfile stayed the same.
Video: Overview of Docker Layer Caching
In the video example, the job runs all of the steps in a Dockerfile with the docker_layer_caching: true for the setup_remote_docker step. On subsequent runs of that job, steps that haven’t changed in the Dockerfile, will be reused. So, the first run takes over two minutes to build the Docker image. If nothing changes in the Dockerfile before the second run, those steps happen instantly, in zero seconds.
version: 2
jobs:
build:
docker:
- image: circleci/node:9.8.0-stretch-browsers
steps:
- checkout
- setup_remote_docker:
docker_layer_caching: true
- run: docker build .
When none of the layers in the image change between job runs, DLC pulls the layers from cache from the image that was built previously and reuses those instead of rebuilding the entire image.
If part of the Dockerfile changes (which changes part of the image) a subsequent run of the exact same job with the modified Dockerfile may still finish faster than rebuilding the entire image. It will finish faster because the cache is used for the first few steps that didn’t change in the Dockerfile. The steps that follow the change must be rerun because the Dockerfile change invalidates the cache.
So, if you change something in the Dockerfile, all of those later steps are invalidated and the layers have to be rebuilt. When some of the steps remain the same (the steps before the one you removed), those steps can be reused. So, it is still faster than rebuilding the entire image.