Home

Awesome

dbt-beyond-the-basics

CI License: MIT

A repository demonstrating advanced use cases of dbt in the following areas:

See something incorrect, open an issue!

Want to see something else included, open an issue 😉!

Continuous Integration

Continuous Integration (CI) is the process of codifying standards, these range from formatting of file contents to validating the correctness of generated data in a data warehouse.

Pre-commit

Pre-commit provides a standardised process to run CI before committing to your local branch. This has several benefits, primarily providing the developer with a quick feedback loop on their work as well as ensuring changes that do not align with standards are automatically identified before being merged. Pre-commit operates via hooks, all of these hooks are sepecified in a .pre-commit-config.yamlfile. There are several hooks that are relevant to a dbt project:

The advantage of local hooks

Most pre-commit hooks are "isolated" hooks in the sense that pre-commit creates a dedicated, isolated environment for each hook to run in. In effect this means that the python environment the hook runs in is not the same as the python environment you are working in locally.

For example, you pip install the sqlfmt package and your local environment now has version 0.23.0 installed. You may run sqlfmt models to format your dbt models after making some changes. When you are ready to commit your changes pre-commit also runs sqlfmt, however it will use a different python environment to do so, potentially resulting in conflicting changes.

One way to avoid this is to use local hooks. These are hooks that run in the same python environment that you are developing in. For example, this "isolated" hook:

# .pre-commit-config.yaml
- repo: https://github.com/tconbeer/sqlfmt
    rev: v0.24.0
    hooks:
    - id: sqlfmt

Can be changed to:

# .pre-commit-config.yaml
- repo: local
hooks:
    - id: sqlfmt
    entry: python -m sqlfmt
    language: system
    name: Run sqlfmt
    pass_filenames: true
    types_or: [jinja, sql]

The primary advantage of this change is that your local environment and pre-commit are now configured to use the same python environment and the same sqlfmt version. A tangential benefit is that updates to packages used in pre-commit now only require updating of the python package. Previously this would have required updating both the python package and the pre-commit hook, a process which if not done correctly could result in a mis-matched setup.

dbt Artifacts and Pytest

dbt produces 4 artifacts in the form of JSON files:

All artifacts are saved in the ./target directory by default.

These JSON files provide a valuable resource when it comes to understanding our dbt project and codifying standards. To run tests on these these files we use pytest, a python based testing framework:

The most valuable artifacts for this are catalog.json and manifest.json. Example tests include:

These tests can (and should) be run in the CI pipeline:

# ./.github/workflows/ci_pipeline.yml
- run: pytest ./tests/pytest -m no_deps

They can also be run as a pre-commit hook:

# .pre-commit-config.yaml
- repo: local
    hooks:
    - id: pytest-catalog-json
        name: pytest-catalog-json
        entry: pytest ./tests/pytest -m catalog_json
        language: system
        pass_filenames: false
        always_run: true

Coverage reports

Some of the functionality discussed above in dbt Artifacts and Pytest can be automated using dbt-coverage. This is a python package that prduces coverage reports for both documentation and, separately, for tests. All pull requests in this repo will have a comment that provides these stats. This allows PR reviewers to quickly assess if any newly added models are lacking acceptable documentation or test coverage.

dbt-bouncer

As an alternative to running pytest in our CI pipeline we can instead use dbt-bouncer. This is a python package that runs a series of checks on a dbt project.

Running dbt-bouncer involves three steps:

  1. Install the package:

    pip install dbt-bouncer

  2. Create a dbt-bouncer.yml configuration file, see dbt-bouncer.yml for an example. This file lists all the checks we want to apply to this dbt project.

  3. Run the dbt-bouncer command (locally or in a CI pipeline):

    dbt-bouncer

dbt commands

Any CI pipeline should run several dbt commands:

All build commands should make use of the following flags:

An example dbt build command as part of the CI pipeline:

# ./.github/workflows/ci_pipeline.yml
- run: dbt --warn-error build --fail-fast

Using state:modified

As part of the CI pipeline the manifest.json artifact is generated for the feature branch, this can be compared to the manifest.json of the target branch using the state method to identify any nodes that have been modified. In addition, the use of the state:modified+ flag allows all downstream nodes to also be identified. When combined with exposures and comments in the PR this can help reviewers quickly assess the potential impact of a PR.

PR comment showing modified nodes and downstream exposures

Mart Monitor

A popular approach to CI for dbt is running Slim CI, this runs the modified nodes and all downstream nodes. This has the benefit of only testing modified nodes and therefore reducing run times and operational costs.

In certain setups it may be desireable to run the entire dbt project in every CI pipeline run. While this sounds extreme there are several methods that can be used to retain the benefits of Slim CI while benefiting from other advantages, namely the ability to provide comprehensive feedback on the impact of a PR on mart models. This can be performed via several steps:

A mart monitor that needs to be investigated further, source.

A mart monitor indicating a mart model has not changed, source.

A downside of building all models in a CI pipeline is increased run time and resource consumption. This can be restricted via pytests based on the run_results.json artifact. See ./tests/pytest/run_results.py for examples of how the duration and resource consumption of dbt build in the CI pipeline can be set to have reasonable allowable values. This provides a number of benefits:

Continuous Deployment

dbt Docs

dbt Docs is a static website that exposes all documentation relating to your dbt project. Normally this is generated and served locally via:

dbt docs generate
dbt docs serve

This works well for the dbt developer as their local python environment is already set up to support these commands. But this isn't an option for some data consumers like the head of marketing who wants to understand what a metric means or a financial analysts looking for the most suitable table to query. For these data consumers we can expose the dbt Docs website via a web server, in our case we can use GitHub Pages.

Every time we push to our prd branch, the cd_dbt_docs.yml workflow is triggered. This workflow runs the above dbt commands and uses the peaceiris/actions-gh-pages GitHub Action to expose the generated dbt Docs website.

GitHub Pages is awesome as it is free for personal, public repositories (like this repository) and also for organisations with an Enterprise plan. If your organisation has GitHub Pages, these are placed behind the same SSO as your GitHub repositories, providing a safe way of exposing dbt Docs to members of your organisation. If you do not use GitHub, there are many alternatives available such as Cloudflare and Netlify, in addition AWS, Azure and GCP can all serve static websites from their cloud storage products.

Entity Relationship Diagram (ERD)

An entity relationship diagram (ERD) is a visual representation of the relationships between entities in a database. It is a useful tool for understanding the structure of a database and can be used by dbt developers when adding new features and also by analysts when writing queries to answer business questions. For a dbt project, only the marts layer is exposed to end users, hence the ERD only needs to include this layer. Here is the ERD for this dbt project (it's rather basic, a real-world dbt project would have a significantly busier ERD):

How is this created?

  1. In our marts layer we have defined relationships tests. For example, the customer_id column in dim_customers is related to the customer_id column in dim_orders.

  2. Using dbterd we can generate a mermaid diagram of the marts layer including the relationships.

  3. Using mermaid-py we can convert the mermaid diagram to a png image.

  4. Combining the last two steps into a single python script: ./scripts/generate_marts_erd_diagram.py.

  5. Now the tricky part. I want this diagram in the README of my repository. But I don't want every developer to have to run this script before creating their PR, I want the image to be automatically generated and kept up to date. So I do the following:

    1. Using the cd_erd_diagram.yml workflow I trigger a GitHub workflow after every merge to prd.

    2. This workflow runs the script and generates the ERD image.

    3. The image is then force-pushed to the erd-diagram branch. Why? The prd branch has branch protection rules that prevent force pushes (this is good practice), so pushing to a different branch avoids this.

    4. I now have an automatically updated ERD available at a static URL:

      https://github.com/pgoslatara/dbt-beyond-the-basics/blob/erd-diagram/target/mermaid.png?raw=true

      This is the URL I reference for the above image.

Docker-ising dbt

"Docker-ising" refers to building a Docker image that can run your dbt project. This is useful when your orchestration process involves running a command in a pre-built container, think Airflow, Cloud Build, Dagster or any other modern Cloud orchestration tool. When building a Docker image for dbt we want to follow several guidelines:

In CI, the ci_pipeline.yml workflow builds a Docker image and runs a dbt parse command on it to validate that the image can run dbt commands.

In CD, although not implemented in this repository, the docker/build-push-action GitHub Action can be used to push the image to an image registry such as GCP's Artifact Registry or AWS's ECR. From here the image can be downloaded by your orchestration tool. It is common to tag images with the SHA of the commit that built the image, in addition you can tag images with the environment they are intended to be used in. For example, an image will initially have the commit SHA and stg tag, it will then be used in our staging environment, after a deployment the stg tag will be replaced with a prd tag and the image will be used in production. When a subsequent deployment to production is performed the prd tag is re-assigned to a newer image and the original image retains only its SHA tag. At all times there is one image with a stg tag and one image with a prd tag. Immediately after a deployment from staging to production, one image will have both tags (i.e. staging and production will use the same image).

Dev Containers

Dev containers provide a Docker-ised development environment and are natively supported by both PyCharm and VSCode, allowing developers to continue using their preferred IDE. Using a dev container allows all developers to work in a standardised environment (including VSCode extensions!), minimising setup issues, reducing the need for manual configuration and allowing for a consistent development experience. Dev containers are useful in workplaces where developers use different OS's (think Mac and Windows), where developers may not be familiar with setting up python environments and where connecting to the underlying database requires non-standard configuration (SQL Server sometimes requires specific drivers to be installed). You can even use the devcontainers/ci Github Action to use your standardised dev container in your GitHub workflows.

To use a dev container you must have Docker (or another container manager like Podman) installed, while some workplaces may restrict this due to security concerns, container managers are very widely used engineering tools and when used correctly can be used for more than just dev containers.

To view the dev container configuration for this project, view the .devcontainer directory.

To open this repository in a dev container:

  1. Click this button: Open in dev container

  2. Clone this repo to your local machine, open the repository in VSCode and from the command palette select Dev Containers: Reopen in Container.

Python

dbt runs in a python environment, therefore the configuration of your python environment is a critical part of a dbt project.

The .python-version file

There are many different versions of python, and there are many different parts of a dbt project that require access to python. One widely supported way of managing the python version is to create a .python-version file in the root of your project. This file contains the python version you want to use, and is as simple as:

3.11.10

Python has a large ecosystem of tools, many of these will use the .python-version file if it is present:

Package Managers

This repository uses Poetry as a python package manager. Package managers are used to install and manage python packages, one of their primary benefits is the generation of a lock file, a file detailing the exact version of every installed package. For Poetry, this is the poetry.lock file. This helps control what are known as transitive dependencies, dependencies that are installed as a result of installing a package. For example, if I was using pip to install packages I may specify dbt-core>=1.8,<1.9. With Poetry I would specify dbt-core=">=1.8.0,<1.9.0". The installed version of dbt-core will be the same using both methods. However dbt-core has dependencies (such as click, logbook, etc.). With pip I have no control over the version of these dependencies, with Poetry I do as the lock file ensures even these dependencies are recorded.

Note that there are several other python package managers available such as hatch, pdm and uv.

Caching in GitHub Workflows

GitHub workflows initialise in an almost empty environment, just your repository contents and some standard CLI tools (think gh, GitHub's own CLI tool). One common step is to recreate the python virtual environment in the workflow environment. This involves downloading and installing all the python dependencies into the .venv directory. But what if you are re-running a CI workflow, or what if your dependencies are the same as your previous run? In these cases you are repeating work that has already been done.

One way to avoid this is to use caching, the concept of storing the output of your work and reusing it in a later run. GitHub Actions has a built in caching mechanism, you can specify a cache key and a path to cache. You can even view the existing caches for this repo here.

dbt repositories can benefit from caching in two areas:

  1. The python virtual environment stored in .venv.
  2. The Poetry executable stored in /home/runner/.local.

Both of these are the result of work we perform in almost every GitHub workflow run. And both of these can be easily invalidated when necessary; the virtual environment when the contents of poetry.lock change and the Poetry executable when the version of Poetry changes.

To enable caching, let's take an example workflow snippet:


env:
  POETRY_VERSION: "1.8.3"

jobs:
  auto-update:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-python@v5
        id: setup-python

      - name: Load cached Poetry installation
        id: cached-poetry
        uses: actions/cache@v4
        with:
              path: /home/runner/.local
              key: poetry-cache-${{ runner.os }}-${{ steps.setup-python.outputs.python-version }}-${{ env.POETRY_VERSION }}

      - name: Install Poetry
        if: steps.cached-poetry.outputs.cache-hit != 'true'
        uses: snok/install-poetry@v1
        with:
              installer-parallel: true
              version: ${{ env.POETRY_VERSION }}
              virtualenvs-create: false
              virtualenvs-in-project: true

      - name: Load cached venv
        id: cached-poetry-dependencies
        uses: actions/cache@v4
        with:
              path: .venv
              key: venv-${{ runner.os }}-${{ steps.setup-python.outputs.python-version }}-${{ hashFiles('**/poetry.lock') }}

      - name: Install python packages
        if: steps.cached-poetry-dependencies.outputs.cache-hit != 'true'
        run: poetry install --no-interaction --no-ansi

      - name: Whatever else we want to do
        run: ...

Every cache requires a unique identifier, this key should contain a reference to the parameters the cache is dependent upon. For the Poetry cache, this is the Poetry version and the python version and the operating system. For the virtual environment cache, this is the operating system, the python version and the hash of the poetry.lock file. By combining these parameters we create a key that allows us to check if a suitable cache already exists, and if not, create a new one.

If a cache is found, then it is loaded. We can this skip subsequent steps like installing Poetry by adding if condition to the step that would have performed that work.

Others

Running dbt from python

In version 1.5, dbt introduced programmatic invocations, a way of calling dbt commands natively from python including the ability to retrieve returned data. Previous ways of doing this mostly relied on opening a new shell process and calling the dbt CLI, this wasn't ideal for a lot of reasons including security. This repo further abstracts programmatic invocations to a dedicated helper function, see run_dbt_command in ./scripts/utils.py.

Conferences

This repository accompanies some conference talks: