Cache dependencies in GitLab CI/CD

GitLab CI/CD provides a caching mechanism that can be used to save time when your jobs are running.

Caching is about speeding the time a job is executed by reusing the same content of a previous job. It can be particularly useful when you are developing software that depends on other libraries which are fetched via the internet during build time.

If caching is enabled, it's shared between pipelines and jobs at the project level by default, starting from GitLab 9.0. Caches are not shared across projects.

Make sure you read the cache reference to learn how it is defined in .gitlab-ci.yml.

Cache vs artifacts

NOTE: Note: Be careful if you use cache and artifacts to store the same path in your jobs as caches are restored before artifacts and the content could be overwritten.

Don't use caching for passing artifacts between stages, as it is designed to store runtime dependencies needed to compile the project:

cache: For storing project dependencies

Caches are used to speed up runs of a given job in subsequent pipelines, by storing downloaded dependencies so that they don't have to be fetched from the internet again (like npm packages, Go vendor packages, etc.) While the cache could be configured to pass intermediate build results between stages, this should be done with artifacts instead.
artifacts: Use for stage results that will be passed between stages.

Artifacts are files generated by a job which are stored and uploaded, and can then be fetched and used by jobs in later stages of the same pipeline. In other words, you can't create an artifact in job-A in stage-1, and then use this artifact in job-B in stage-1. This data will not be available in different pipelines, but is available to be downloaded from the UI.

The name artifacts sounds like it's only useful outside of the job, like for downloading a final image, but artifacts are also available in later stages within a pipeline. So if you build your application by downloading all the required modules, you might want to declare them as artifacts so that subsequent stages can use them. There are some optimizations like declaring an expiry time so you don't keep artifacts around too long, or using dependencies to control which jobs fetch the artifacts.

Caches:

Are disabled if not defined globally or per job (using cache:).
Are available for all jobs in your .gitlab-ci.yml if enabled globally.
Can be used in subsequent pipelines by the same job in which the cache was created (if not defined globally).
Are stored where GitLab Runner is installed and uploaded to S3 if distributed cache is enabled.
If defined per job, are used:
- By the same job in a subsequent pipeline.
- By subsequent jobs in the same pipeline, if they have identical dependencies.

Artifacts:

Are disabled if not defined per job (using artifacts:).
Can only be enabled per job, not globally.
Are created during a pipeline and can be used by the subsequent jobs of that currently active pipeline.
Are always uploaded to GitLab (known as coordinator).
Can have an expiration value for controlling disk usage (30 days by default).

NOTE: Note: Both artifacts and caches define their paths relative to the project directory, and can't link to files outside it.

Good caching practices

We have the cache from the perspective of the developers (who consume a cache within the job) and the cache from the perspective of the runner. Depending on which type of runner you are using, cache can act differently.

From the perspective of the developer, to ensure maximum availability of the cache, when declaring cache in your jobs, use one or a mix of the following:

Tag your runners and use the tag on jobs that share their cache.
Use sticky runners that will be only available to a particular project.
Use a key that fits your workflow (for example, different caches on each branch). For that, you can take advantage of the CI/CD predefined variables.

TIP: Tip: Using the same runner for your pipeline, is the most simple and efficient way to cache files in one stage or pipeline, and pass this cache to subsequent stages or pipelines in a guaranteed manner.

From the perspective of the runner, in order for cache to work effectively, one of the following must be true:

Use a single runner for all your jobs.
Use multiple runners (in autoscale mode or not) that use distributed caching, where the cache is stored in S3 buckets (like shared runners on GitLab.com).
Use multiple runners (not in autoscale mode) of the same architecture that share a common network-mounted directory (using NFS or something similar) where the cache will be stored.

TIP: Tip: Read about the availability of the cache to learn more about the internals and get a better idea how cache works.

Sharing caches across the same branch

Define a cache with the key: ${CI_COMMIT_REF_SLUG} so that jobs of each branch always use the same cache:

cache:
  key: ${CI_COMMIT_REF_SLUG}

While this feels like it might be safe from accidentally overwriting the cache, it means merge requests get slow first pipelines, which might be a bad developer experience. The next time a new commit is pushed to the branch, the cache will be re-used.

To enable per-job and per-branch caching:

cache:
  key: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"

To enable per-branch and per-stage caching:

cache:
  key: "$CI_JOB_STAGE-$CI_COMMIT_REF_SLUG"

Sharing caches across different branches

If the files you are caching need to be shared across all branches and all jobs, you can use the same key for all of them:

cache:
  key: one-key-to-rule-them-all

To share the same cache between branches, but separate them by job:

cache:
  key: ${CI_JOB_NAME}

Disabling cache on specific jobs

If you have defined the cache globally, it means that each job will use the same definition. You can override this behavior per-job, and if you want to disable it completely, use an empty hash:

job:
  cache: {}

Inherit global config, but override specific settings per job

You can override cache settings without overwriting the global cache by using anchors. For example, if you want to override the policy for one job:

cache: &global_cache
  key: ${CI_COMMIT_REF_SLUG}
  paths:
    - node_modules/
    - public/
    - vendor/
  policy: pull-push

job:
  cache:
    # inherit all global cache settings
    <<: *global_cache
    # override the policy
    policy: pull

For more fine tuning, read also about the cache: policy.

Common use cases

The most common use case of cache is to preserve contents between subsequent runs of jobs for things like dependencies and commonly used libraries (Node.js packages, PHP packages, rubygems, Python libraries, etc.), so they don't have to be re-fetched from the public internet.

NOTE: Note: For more examples, check out our GitLab CI/CD templates.

Caching Node.js dependencies

Assuming your project is using npm to install the Node.js dependencies, the following example defines cache globally so that all jobs inherit it. By default, npm stores cache data in the home folder ~/.npm but since you can't cache things outside of the project directory, we tell npm to use ./.npm instead, and it is cached per-branch:

#
# https://gitlab.com/gitlab-org/gitlab/tree/master/lib/gitlab/ci/templates/Nodejs.gitlab-ci.yml
#
image: node:latest

# Cache modules in between jobs
cache:
  key: ${CI_COMMIT_REF_SLUG}
  paths:
    - .npm/

before_script:
  - npm ci --cache .npm --prefer-offline

test_async:
  script:
    - node ./specs/start.js ./specs/async.spec.js

Caching PHP dependencies

Assuming your project is using Composer to install the PHP dependencies, the following example defines cache globally so that all jobs inherit it. PHP libraries modules are installed in vendor/ and are cached per-branch:

#
# https://gitlab.com/gitlab-org/gitlab/tree/master/lib/gitlab/ci/templates/PHP.gitlab-ci.yml
#
image: php:7.2

# Cache libraries in between jobs
cache:
  key: ${CI_COMMIT_REF_SLUG}
  paths:
    - vendor/

before_script:
  # Install and run Composer
  - curl --show-error --silent https://getcomposer.org/installer | php
  - php composer.phar install

test:
  script:
    - vendor/bin/phpunit --configuration phpunit.xml --coverage-text --colors=never

Caching Python dependencies

Assuming your project is using pip to install the Python dependencies, the following example defines cache globally so that all jobs inherit it. Python libraries are installed in a virtual environment under venv/, pip's cache is defined under .cache/pip/ and both are cached per-branch:

#
# https://gitlab.com/gitlab-org/gitlab/tree/master/lib/gitlab/ci/templates/Python.gitlab-ci.yml
#
image: python:latest

# Change pip's cache directory to be inside the project directory since we can
# only cache local items.
variables:
  PIP_CACHE_DIR: "$CI_PROJECT_DIR/.cache/pip"

# Pip's cache doesn't store the python packages
# https://pip.pypa.io/en/stable/reference/pip_install/#caching
#
# If you want to also cache the installed packages, you have to install
# them in a virtualenv and cache it as well.
cache:
  paths:
    - .cache/pip
    - venv/

before_script:
  - python -V               # Print out python version for debugging
  - pip install virtualenv
  - virtualenv venv
  - source venv/bin/activate

test:
  script:
    - python setup.py test
    - pip install flake8
    - flake8 .

Caching Ruby dependencies

Assuming your project is using Bundler to install the gem dependencies, the following example defines cache globally so that all jobs inherit it. Gems are installed in vendor/ruby/ and are cached per-branch:

#
# https://gitlab.com/gitlab-org/gitlab/tree/master/lib/gitlab/ci/templates/Ruby.gitlab-ci.yml
#
image: ruby:2.6

# Cache gems in between builds
cache:
  key: ${CI_COMMIT_REF_SLUG}
  paths:
    - vendor/ruby

before_script:
  - ruby -v                                        # Print out ruby version for debugging
  - bundle install -j $(nproc) --path vendor/ruby  # Install dependencies into ./vendor/ruby

rspec:
  script:
    - rspec spec

Caching Go dependencies

Assuming your project is using Go Modules to install Go dependencies, the following example defines cache in a go-cache template, that any job can extend. Go modules are installed in ${GOPATH}/pkg/mod/ and are cached for all of the go projects:

.go-cache:
  variables:
    GOPATH: $CI_PROJECT_DIR/.go
  before_script:
    - mkdir -p .go
  cache:
    paths:
      - .go/pkg/mod/

test:
  image: golang:1.13
  extends: .go-cache
  script:
    - go test ./... -v -short

Availability of the cache

Caching is an optimization, but isn't guaranteed to always work, so you need to be prepared to regenerate any cached files in each job that needs them.

Assuming you have properly defined cache in .gitlab-ci.yml according to your workflow, the availability of the cache ultimately depends on how the runner has been configured (the executor type and whether different runners are used for passing the cache between jobs).

Where the caches are stored

Since the runner is the one responsible for storing the cache, it's essential to know where it's stored. All the cache paths defined under a job in .gitlab-ci.yml are archived in a single cache.zip file and stored in the runner's configured cache location. By default, they are stored locally in the machine where the runner is installed and depends on the type of the executor.

GitLab Runner executor	Default path of the cache
Shell	Locally, stored under the `gitlab-runner` user's home directory: `/home/gitlab-runner/cache/<user>/<project>/<cache-key>/cache.zip`.
Docker	Locally, stored under Docker volumes: `/var/lib/docker/volumes/<volume-id>/_data/<user>/<project>/<cache-key>/cache.zip`.
Docker machine (autoscale runners)	Behaves the same as the Docker executor.

How archiving and extracting works

In the most simple scenario, consider that you use only one machine where the runner is installed, and all jobs of your project run on the same host.

Let's see the following example of two jobs that belong to two consecutive stages:

stages:
  - build
  - test

before_script:
  - echo "Hello"

job A:
  stage: build
  script:
    - mkdir vendor/
    - echo "build" > vendor/hello.txt
  cache:
    key: build-cache
    paths:
      - vendor/
  after_script:
    - echo "World"

job B:
  stage: test
  script:
    - cat vendor/hello.txt
  cache:
    key: build-cache

Here's what happens behind the scenes:

Pipeline starts.
job A runs.
before_script is executed.
script is executed.
after_script is executed.
cache runs and the vendor/ directory is zipped into cache.zip. This file is then saved in the directory based on the runner's setting and the cache: key.
job B runs.
The cache is extracted (if found).
before_script is executed.
script is executed.
Pipeline finishes.

By using a single runner on a single machine, you'll not have the issue where job B might execute on a runner different from job A, thus guaranteeing the cache between stages. That will only work if the build goes from stage build to test in the same runner/machine, otherwise, you might not have the cache available.

During the caching process, there's also a couple of things to consider:

If some other job, with another cache configuration had saved its cache in the same zip file, it is overwritten. If the S3 based shared cache is used, the file is additionally uploaded to S3 to an object based on the cache key. So, two jobs with different paths, but the same cache key, will overwrite their cache.
When extracting the cache from cache.zip, everything in the zip file is extracted in the job's working directory (usually the repository which is pulled down), and the runner doesn't mind if the archive of job A overwrites things in the archive of job B.

The reason why it works this way is because the cache created for one runner often will not be valid when used by a different one which can run on a different architecture (e.g., when the cache includes binary files). And since the different steps might be executed by runners running on different machines, it is a safe default.

Cache mismatch

In the following table, you can see some reasons where you might hit a cache mismatch and a few ideas how to fix it.

Reason of a cache mismatch	How to fix it
You use multiple standalone runners (not in autoscale mode) attached to one project without a shared cache	Use only one runner for your project or use multiple runners with distributed cache enabled
You use runners in autoscale mode without a distributed cache enabled	Configure the autoscale runner to use a distributed cache
The machine the runner is installed on is low on disk space or, if you've set up distributed cache, the S3 bucket where the cache is stored doesn't have enough space	Make sure you clear some space to allow new caches to be stored. Currently, there's no automatic way to do this.
You use the same `key` for jobs where they cache different paths.	Use different cache keys to that the cache archive is stored to a different location and doesn't overwrite wrong caches.

Let's explore some examples.

Examples

Let's assume you have only one runner assigned to your project, so the cache will be stored in the runner's machine by default. If two jobs, A and B, have the same cache key, but they cache different paths, cache B would overwrite cache A, even if their paths don't match:

We want job A and job B to re-use their cache when the pipeline is run for a second time.

stages:
  - build
  - test

job A:
  stage: build
  script: make build
  cache:
    key: same-key
    paths:
      - public/

job B:
  stage: test
  script: make test
  cache:
    key: same-key
    paths:
      - vendor/

job A runs.
public/ is cached as cache.zip.
job B runs.
The previous cache, if any, is unzipped.
vendor/ is cached as cache.zip and overwrites the previous one.
The next time job A runs it will use the cache of job B which is different and thus will be ineffective.

To fix that, use different keys for each job.

In another case, let's assume you have more than one runner assigned to your project, but the distributed cache is not enabled. The second time the pipeline is run, we want job A and job B to re-use their cache (which in this case will be different):

stages:
  - build
  - test

job A:
  stage: build
  script: build
  cache:
    key: keyA
    paths:
      - vendor/

job B:
  stage: test
  script: test
  cache:
    key: keyB
    paths:
      - vendor/

In that case, even if the key is different (no fear of overwriting), you might experience that the cached files "get cleaned" before each stage if the jobs run on different runners in the subsequent pipelines.

Clearing the cache

Runners use cache to speed up the execution of your jobs by reusing existing data. This however, can sometimes lead to an inconsistent behavior.

To start with a fresh copy of the cache, there are two ways to do that.

Clearing the cache by changing `cache:key`

All you have to do is set a new cache: key in your .gitlab-ci.yml. In the next run of the pipeline, the cache will be stored in a different location.

Clearing the cache manually

Introduced in GitLab 10.4.

If you want to avoid editing .gitlab-ci.yml, you can easily clear the cache via GitLab's UI:

Navigate to your project's CI/CD > Pipelines page.
Click on the Clear runner caches button to clean up the cache.
On the next push, your CI/CD job will use a new cache.

Behind the scenes, this works by increasing a counter in the database, and the value of that counter is used to create the key for the cache by appending an integer to it: -1, -2, etc. After a push, a new key is generated and the old cache is not valid anymore.