Caching in GitLab CI/CD

A cache is one or more files that a job downloads and saves. Subsequent jobs that use the same cache don’t have to download the files again, so they execute more quickly.

To learn how to define the cache in your .gitlab-ci.yml file, see the cache reference.

How cache is different from artifacts

Use cache for dependencies, like packages you download from the internet. Cache is stored where GitLab Runner is installed and uploaded to S3 if distributed cache is enabled.

Use artifacts to pass intermediate build results between stages. Artifacts are generated by a job, stored in GitLab, and can be downloaded.

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

Cache

  • Define cache per job by using the cache: keyword. Otherwise it is disabled.
  • Subsequent pipelines can use the cache.
  • Subsequent jobs in the same pipeline can use the cache, if the dependencies are identical.
  • Different projects cannot share the cache.

Artifacts

  • Define artifacts per job.
  • Subsequent jobs in later stages of the same pipeline can use artifacts.
  • Different projects cannot share artifacts.

Artifacts expire after 30 days unless you define an expiration time. Use dependencies to control which jobs fetch the artifacts.

Good caching practices

To ensure maximum availability of the cache, do one or more of the following:

For runners to work with caches efficiently, you must do one of the following:

  • Use a single runner for all your jobs.
  • Use multiple runners that have distributed caching, where the cache is stored in S3 buckets. Shared runners on GitLab.com behave this way. These runners can be in autoscale mode, but they don’t have to be.
  • Use multiple runners with the same architecture and have these runners share a common network-mounted directory to store the cache. This directory should use NFS or something similar. These runners must be in autoscale mode.

Use multiple caches

Version history

You can have a maximum of four caches:

test-job:
  stage: build
  cache:
    - key:
        files:
          - Gemfile.lock
      paths:
        - vendor/ruby
    - key:
        files:
          - yarn.lock
      paths:
        - .yarn-cache/
  script:
    - bundle install --path=vendor
    - yarn install --cache-folder .yarn-cache
    - echo Run tests...

If multiple caches are combined with a Fallback cache key, the fallback cache is fetched every time a cache is not found.

Fallback cache key

Introduced in GitLab Runner 13.4.

You can use the $CI_COMMIT_REF_SLUG predefined variable to specify your cache:key. For example, if your $CI_COMMIT_REF_SLUG is test you can set a job to download cache that’s tagged with test.

If a cache with this tag is not found, you can use CACHE_FALLBACK_KEY to specify a cache to use when none exists.

In the following example, if the $CI_COMMIT_REF_SLUG is not found, the job uses the key defined by the CACHE_FALLBACK_KEY variable:

variables:
  CACHE_FALLBACK_KEY: fallback-key

job1:
  script:
    - echo
  cache:
    key: "$CI_COMMIT_REF_SLUG"
    paths:
      - binaries/

Disable cache for specific jobs

If you have defined the cache globally, it means that each job uses 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 configuration, 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 caching is to avoid downloading content like dependencies or libraries repeatedly between subsequent runs of jobs. Node.js packages, PHP packages, Ruby gems, Python libraries, and others can all be cached.

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

Share caches between jobs in the same branch

To have jobs for each branch use the same cache, define a cache with the key: $CI_COMMIT_REF_SLUG:

cache:
  key: $CI_COMMIT_REF_SLUG

This configuration prevents you from accidentally overwriting the cache. However, the first pipeline for a merge request is slow. The next time a commit is pushed to the branch, the cache is re-used and jobs run faster.

To enable per-job and per-branch caching:

cache:
  key: "$CI_JOB_NAME-$CI_COMMIT_REF_SLUG"

To enable per-stage and per-branch caching:

cache:
  key: "$CI_JOB_STAGE-$CI_COMMIT_REF_SLUG"

Share caches across jobs in different branches

To share a cache across all branches and all jobs, use the same key for everything:

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

To share caches between branches, but have a unique cache for each job:

cache:
  key: $CI_JOB_NAME

Cache Node.js dependencies

If 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 you can’t cache things outside of the project directory. Instead, we tell npm to use ./.npm, and cache it 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

Cache 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

Cache 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 .

Cache 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

If you have jobs that each need a different selection of gems, use the prefix keyword in the global cache definition. This configuration generates a different cache for each job.

For example, a testing job might not need the same gems as a job that deploys to production:

cache:
  key:
    files:
      - Gemfile.lock
    prefix: $CI_JOB_NAME
  paths:
    - vendor/ruby

test_job:
  stage: test
  before_script:
    - bundle install --without production --path vendor/ruby
  script:
    - bundle exec rspec

deploy_job:
  stage: production
  before_script:
    - bundle install --without test --path vendor/ruby
  script:
    - bundle exec deploy

Cache 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 it isn’t guaranteed to always work. You need to be prepared to regenerate any cached files in each job that needs them.

After you have defined a cache in .gitlab-ci.yml, the availability of the cache depends on:

  • The runner’s executor type
  • Whether different runners are used to pass the cache between jobs.

Where the caches are stored

The runner is responsible for storing the cache, so 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.

If you use cache and artifacts to store the same path in your jobs, the cache might be overwritten because caches are restored before artifacts.

How archiving and extracting works

This example has 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
    paths:
      - vendor/

If you have one machine with one runner installed, and all jobs for your project run on the same host:

  1. Pipeline starts.
  2. job A runs.
  3. before_script is executed.
  4. script is executed.
  5. after_script is executed.
  6. 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.
  7. job B runs.
  8. The cache is extracted (if found).
  9. before_script is executed.
  10. script is executed.
  11. Pipeline finishes.

By using a single runner on a single machine, you don’t have the issue where job B might execute on a runner different from job A. This setup guarantees the cache can be reused between stages. It only works if the execution goes from the build stage to the test stage in the same runner/machine. Otherwise, the cache might not be 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, overwrites 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.

It works this way because the cache created for one runner often isn’t valid when used by a different one. A different runner may run on a different architecture (for example, when the cache includes binary files). Also, because 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. 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 is stored in the runner’s machine by default.

Two jobs could cause caches to be overwritten if they have the same cache key, but they cache a different path:

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/
  1. job A runs.
  2. public/ is cached as cache.zip.
  3. job B runs.
  4. The previous cache, if any, is unzipped.
  5. vendor/ is cached as cache.zip and overwrites the previous one.
  6. The next time job A runs it uses the cache of job B which is different and thus isn’t effective.

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 is 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/

Even if the key is different, the cached files might 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 is 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 clear the cache via the GitLab UI:

  1. Navigate to your project’s CI/CD > Pipelines page.
  2. Click on the Clear runner caches button to clean up the cache.

    Clear runner caches

  3. On the next push, your CI/CD job uses a new cache.
noteEach time you clear the cache manually, the internal cache name is updated. The name uses the format cache-<index>, and the index increments by one each time. The old cache is not deleted. You can manually delete these files from the runner storage.