Skip to main content
Version: Next

Docker builds, images and tags

The Apache Superset community extensively uses Docker for development, release, and productionizing Superset. This page details our Docker builds and tag naming schemes to help users navigate our offerings.

Images are built and pushed to the Superset Docker Hub repository using GitHub Actions. Different sets of images are built and/or published at different times:

  • Published releases (release): published using tags like 5.0.0 and the latest tag.
  • Pull request iterations (pull_request): for each pull request, while we actively build the docker to validate the build, we do not publish those images for security reasons, we simply docker build --load
  • Merges to the main branch (push): resulting in new SHAs, with tags prefixed with master for the latest master version.

Build presets​

We have a set of build "presets" that each represent a combination of parameters for the build, mostly pointing to either different target layer for the build, and/or base image.

Here are the build presets that are exposed through the supersetbot docker utility:

  • lean: The default Docker image, including both frontend and backend. Tags without a build_preset are lean builds (ie: latest, 5.0.0, 4.1.2, ...). lean builds do not contain database drivers, meaning you need to install your own. That applies to analytics databases AND the metadata database. You'll likely want to layer either mysqlclient or psycopg2-binary depending on the metadata database you choose for your installation, plus the required drivers to connect to your analytics database(s).
  • dev: For development, with a headless browser, dev-related utilities and root access. This includes some commonly used database drivers like mysqlclient, psycopg2-binary and some other used for development/CI
  • py311, e.g., Py311: Similar to lean but with a different Python version (in this example, 3.11).
  • ci: For certain CI workloads.
  • websocket: For Superset clusters supporting advanced features.
  • dockerize: Used by Helm in initContainers to wait for database dependencies to be available.

Key tags examples​

  • latest: The latest official release build
  • latest-dev: the -dev image of the latest official release build, with a headless browser and root access.
  • master: The latest build from the master branch, implicitly the lean build preset
  • master-dev: Similar to master but includes a headless browser and root access.
  • pr-5252: The latest commit in PR 5252.
  • 30948dc401b40982cb7c0dbf6ebbe443b2748c1b-dev: A build for this specific SHA, which could be from a master merge, or release.
  • websocket-latest: The WebSocket image for use in a Superset cluster.

For insights or modifications to the build matrix and tagging conventions, check the supersetbot docker subcommand and the docker.yml GitHub action.

Building your own production Docker image​

Every Superset deployment will require its own set of drivers depending on the data warehouse(s), etc. so we recommend that users build their own Docker image by extending the lean image.

Here's an example Dockerfile that does this. Follow the in-line comments to customize it for your desired Superset version and database drivers. The comments also note that a certain feature flag will have to be enabled in your config file.

You would build the image with docker build -t mysuperset:latest . or docker build -t ourcompanysuperset:5.0.0 .

# change this to apache/superset:5.0.0 or whatever version you want to build from;
# otherwise the default is the latest commit on GitHub master branch
FROM apache/superset:master

USER root

# Set environment variable for Playwright
ENV PLAYWRIGHT_BROWSERS_PATH=/usr/local/share/playwright-browsers

# Install packages using uv into the virtual environment
RUN . /app/.venv/bin/activate && \
    uv pip install \
    # install psycopg2 for using PostgreSQL metadata store - could be a MySQL package if using that backend:
    psycopg2-binary \
    # add the driver(s) for your data warehouse(s), in this example we're showing for Microsoft SQL Server:
    pymssql \
    # package needed for using single-sign on authentication:
    Authlib \
    # openpyxl to be able to upload Excel files
    openpyxl \
    # Pillow for Alerts & Reports to generate PDFs of dashboards
    Pillow \
    # install Playwright for taking screenshots for Alerts & Reports. This assumes the feature flag PLAYWRIGHT_REPORTS_AND_THUMBNAILS is enabled
    # That feature flag will default to True starting in 6.0.0
    # Playwright works only with Chrome.
    # If you are still using Selenium instead of Playwright, you would instead install here the selenium package and a headless browser & webdriver
    playwright \
    && playwright install-deps \
    && PLAYWRIGHT_BROWSERS_PATH=/usr/local/share/playwright-browsers playwright install chromium

# Switch back to the superset user
USER superset

CMD ["/app/docker/entrypoints/run-server.sh"]

Adding translations to a custom image​

The pattern above, a small Dockerfile that just extends FROM apache/superset:..., can't add translations after the fact. By the time an official tag is published, its frontend and backend layers have already had non-English translation files stripped out unless BUILD_TRANSLATIONS was set at build time (see below), and there's no superset/translations source tree left in the final image to compile from.

To get translations into your own image, you need to build from the full Superset source (a clone or fork of this repo) rather than extend a published tag. The most efficient way to do this is to append your customizations as one more stage at the end of the repo's own Dockerfile, so Docker can reuse the cached upstream layers and only rebuild what your stage adds:

# Append this to the end of the repo's Dockerfile
# Keep this tag in sync with the branch/tag of the repo you cloned, so the
# translation files built from source match the keys the runtime expects:
FROM apache/superset:5.0.0 AS my-custom-image
USER root

# Pull the translation files out of the earlier build stages (frontend
# .json in `superset-node`, backend .mo in `python-translation-compiler`).
# Those stages' own cleanup only matches single-character extensions, so
# the source `.po` files can still be present here; strip them explicitly
# so this stage only keeps the compiled translations.
COPY --from=superset-node /app/superset/translations superset/translations
COPY --from=python-translation-compiler /app/translations_mo superset/translations
RUN find superset/translations -name '*.po' -delete

USER superset

Then build with:

docker build --target=my-custom-image --build-arg=BUILD_TRANSLATIONS=true -t mysuperset:5.0.0 .

You can combine this with the database-driver/dependency pattern above by adding your own RUN uv pip install ... step before switching back to USER superset. See issue #35959 for the discussion this pattern came out of, credit to the community for working it out.

Key ARGs in Dockerfile​

  • BUILD_TRANSLATIONS: whether to compile non-English translations into the image. When true, the frontend build converts the *.po files to locale JSON and the backend runs pybabel compile to produce *.mo files; both source *.po files are stripped afterward either way. When false (the default), those compile steps are skipped and only en ships. This only takes effect when building the image from source (docker build against this repo's own Dockerfile); it has no effect on a downstream Dockerfile that just extends an already-published tag, see "Adding translations to a custom image" above. Note that the backend pybabel compile step ignores its exit code, so a .po file with a compile error won't fail the build; check the build logs for pybabel warnings if a locale's backend strings aren't showing up.
  • DEV_MODE: whether to skip the frontend build, this is used by our docker-compose dev setup where we mount the local volume and build using webpack in --watch mode, meaning as you alter the code in the local file system, webpack, from within a docker image used for this purpose, will constantly rebuild the frontend as you go. This ARG enables the initial docker-compose build to take much less time and resources
  • INCLUDE_CHROMIUM: whether to include chromium in the backend build so that it can be used as a headless browser for workloads related to "Alerts & Reports" and thumbnail generation
  • INCLUDE_FIREFOX: same as above, but for firefox
  • PY_VER: specifying the base image for the python backend, we don't recommend altering this setting if you're not working on forwards or backwards compatibility

Caching​

To accelerate builds, we follow Docker best practices and use apache/superset-cache.

About database drivers​

Our docker images come with little to zero database driver support since each environment requires different drivers, and maintaining a build with wide database support would be both challenging (dozens of databases, python drivers, and os dependencies) and inefficient (longer build times, larger images, lower layer cache hit rate, ...).

For production use cases, we recommend that you derive our lean image(s) and add database support for the database you need.

On supporting different platforms (namely arm64 AND amd64)​

Currently all automated builds are multi-platform, supporting both linux/arm64 and linux/amd64. This enables higher level constructs like helm and docker compose to point to these images and effectively be multi-platform as well.

Pull requests and master builds are one-image-per-platform so that they can be parallelized and the build matrix for those is more sparse as we don't need to build every build preset on every platform, and generally can be more selective here. For those builds, we suffix tags with -arm where it applies.

Working with Apple silicon​

Apple's current generation of computers uses ARM-based CPUs, and Docker running on MACs seem to require linux/arm64/v8 (at least one user's M2 was configured in that way). Setting the environment variable DOCKER_DEFAULT_PLATFORM to linux/amd64 seems to function in term of leveraging, and building upon the Superset builds provided here.

export DOCKER_DEFAULT_PLATFORM=linux/amd64

Presumably, linux/arm64/v8 would be more optimized for this generation of chips, but less compatible across the ARM ecosystem.