dbt-beyond-the-basics

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

• Continuous Integration (CI)

  • Pre-commit
  • dbt Artifacts and Pytest
  • Coverage reports
  • dbt-bouncer
  • dbt commands
  • Using state:modified
  • Mart Monitor

• Continuous Deployment (CD)

• Dev Containers

• Others

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:

• Pre-commit itself provides several standard hooks that ensure standard behaviour regarding whitespace control, valid YAML files, no presence of private keys and no unresolved merge conflicts. An interesting hook is no-commit-to-branch, this allows the name of the git branch to be standarised, for example to always start with feature/ or to always include a Jira ticket ID to help with tracking of work items.

  # .pre-commit-config.yaml
  - repo: https://github.com/pre-commit/pre-commit-hooks
      rev: v4.4.0
      hooks:
      - id: trailing-whitespace
      - id: check-merge-conflict
      - id: check-yaml
          args: [--unsafe]
      - id: no-commit-to-branch
          name: JIRA ticket ID in branch
          args: ['--pattern', '^((?![A-Z]+[-][0-9]+[-][\S]+).)*$']

• sqlfmt is the SQL formatter used in the dbt Cloud IDE. It is an opinionated formatter with minimal configuration options making it super easy to setup. It forces all .sql files to a standard SQL format thereby reducing the strain on repo readers by having a single, standard format across the repo. An alternative is SQLFluff, which also has pre-commit hooks.

  # .pre-commit-config.yaml
  - repo: https://github.com/tconbeer/sqlfmt
      rev: v0.18.1
      hooks:
      - id: sqlfmt

• dbt-checkpoint is an awesome pre-commit package with multiple well-documented hooks. Some valuable options include ensuring that every model has a description in a YAML file, naming conventions for models in certain folders and that models have a minimum number of tests.

  # .pre-commit-config.yaml
  - repo: https://github.com/dbt-checkpoint/dbt-checkpoint
      rev: v1.1.0
      hooks:
      - id: dbt-compile
      - id: dbt-docs-generate
      - id: check-model-has-properties-file
          name: Check that all models are listed in a YAML file
      - id: check-model-name-contract
          args: [--pattern, "(base_|stg_).*"]
          files: models/staging/

dbt Artifacts and Pytest

dbt produces 4 artifacts in the form of JSON files:

• catalog.json is produced by dbt docs generate and contains all the information displayed in the docs web UI (primarily model schemas and data types).
• manifest.json is produced by dbt compile and is the main source of information for the project including details on all nodes, the dependencies between these nodes as well as both the raw and compiled SQL that will be run.
• run_results.json is produced by any dbt command that runs a node, e.g. dbt build, dbt run, etc. It contains data on the success of each node, the duration of each node and any data returned by the warehouse (adapter responses).
• sources.json is produced by dbt source freshness, similar to run_results.json it contains data on how long each freshness check takes as well as the success or failure of the check.

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:

• Create a fixture for each artifact:

  # ./tests/pytest/conftest.py
  @pytest.fixture(scope="module")
  def catalog_json() -> dict:
      with Path("./target/catalog.json").open() as f:
          data = json.load(f)
      return data

• Write a pytest that takes a fixture as an input parameter and runs as assert statement:

  # ./tests/pytest/test_columns.py
  @pytest.mark.catalog_json
  def test_column_names_models(catalog_json: dict) -> None:
  
      regex_pattern = "[a-z_0-9]*"
  
      for k, v in catalog_json["nodes"].items():
          for col in v["columns"].keys():
              if col.find(".") <= 0:
                  assert (
                      col == re.compile(regex_pattern).match(col)[0]
                  ), f"Column '{col}' in {k} does not align with the existing naming convention ({regex_pattern})."

  Using the @pytest.mark decorator and creating a pytest.ini file allow us to use marks to group pytests, for example grouping all pytests that use the catalog.json artifact.

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

• A naming convention for columns, e.g. no uppercase characters.
• Each source can only be read by one staging model.
• All columns with a data type of DATE have to end with "_date".
• The ./model/staging directory can only have 1 layer of subdirectories.
• Etc.

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:

• dbt build: This runs and tests all the models, ideally in a dedicated schema (set up via the generate_schema_name macro).
• dbt build --select config.materialized:incremental: This runs and tests all incremental models, this is an important step to ensure any incremental logic does not generate invalid SQL.
• dbt source freshness: This tests the freshness of all sources. The output of this command should be forced to success (via || true) as we are not interested in whether our sources are fresh, we are interested in the generated source.json artifact. See ./test/pytest/test_sources.py for an example of how to identify invalid freshness checks.

All build commands should make use of the following flags:

• --warn-error: Any warning results in a failure. This ensures no warnings enter our production branch as these have a higher likelihood to result in failures in the future or be an unintended consequence of the changes in the PR.
• --fail-fast: Any failed node results in the immediate failure of the command. This provides faster feedback to the developer who is waiting on the results of the CI pipeline.

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.

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:

• Add or edit the generate_schema_name macro to force all models to be built in a single schema when the DBT_CICD_RUN environment variable is true.

  # ./macros/generate_schema_name.sql
  {% macro generate_schema_name(custom_schema_name, node) -%}
  
      {% if env_var('DBT_CICD_RUN', 'false') == 'true' %} {{ env_var('DBT_DATASET') }}
  
      {% elif target.name in ['stg', 'prd'] and env_var('DBT_CICD_RUN', 'false') == 'false' %}
  
          {{ node.config.schema }}
  
      {% else %} {{ default__generate_schema_name(custom_schema_name, node) }}
  
      {%- endif -%}
  
  {%- endmacro %}
  

  This results in a scenario where each CI pipeline run has a dedicated dataset:

  

• For staging models with large volumes of historical data there is no need to process all this data in every CI pipeline run. A jinja "if" condition can be utilised to only use a reasonble volume of data during CI runs:

  # ./models/staging/public_datasets/stg_public_datasets__bitcoin_blocks.sql
  {% if env_var('DBT_CICD_RUN', 'false') == 'true' %}
  
      and timestamp_month >= date_trunc(date_sub(current_date(), interval 1 month), month)
  
  {% endif %}

• In .github/workflows/ci_pipeline, set the required environment variables:

  • Set DBT_CICD_RUN to true.

  • Assemble the value of DBT_DATASET to contain the PR number, run number and sha of the latest commit. This ensures that every run of the pipeline will have a unique schema.

• Add a query to ./scripts/mart_monitor_queries.yml that returns a single row of values. This query can test any model and contain any logic however it is best to start with examing high level summaries of mart models as these are the most critical models in a dbt project.

• In the CI pipeline (.github/workflows/ci_pipeline) run dbt build and run the ./scripts/mart_monitor_commenter.py script passing the required arguments.

• For each mart monitor query a comment will be left in the PR to help developers and reviewers quickly assess the impact of the changes on mart models:

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:

• Poor JOIN logic that takes excessive time to compute will be identified.
• Incorrect or non-use of partitioning to select source data will result in failed CI pipelines.
• As a project grows there is continuous focus on the efficiency of CI runs resulting in a developer mindset that places efficiency higher in the priority list.

Continuous Deployment

TODO

Dev Containers

TODO

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:

• NL dbt meetup: 2nd Edition: "CI for dbt: Beyond the basics!", slides available here.
• MDSFest: "CI for dbt: Beyond the basics!", slides available here, video available here.