This repository provides an example of how to create a Data Pipeline that can be deployed as part of FlowEHR.
Using this example, you can build and deploy data transformation code and test it locally via unit tests. Then, you can test it within your dev environment in Azure. Finally, it can be integrated with your deployment of FlowEHR.
Step 1: Press the "Use this repository" button on Github:
Step 2: Clone this repository using VSCode.
Note that you can use a different IDE as well, but you won't be getting the benefits of using Devcontainers.
Step 3: Replace all occurences of helloworld
with the name of the pipeline you are creating.
Step 4: Reopen the repository in Container.
The following guide assumes you know how to create a Data Pipeline on Databricks and PySpark. It also assumes general familiarity with Azure, Devcontainers, Makefile, Azure Data Factory, and Terraform.
For local development, there is a Devcontainer configuration of this repository. To start using it, start your Devcontainer environment.
To run tests, navigate into helloworld
directory in the Terminal:
cd helloworld
Now you can run unit tests for the Data Pipeline. In the Terminal, run
make test
The source code for tests can be found in ./helloworld/tests/.
You are strongly urged to write individual unit tests for each transform in the data pipeline.
To make sure the Data Pipeline can be deployed by FlowEHR, the pipeline must have two components:
- An
pipeline.json
file. This should conform to the way a pipeline is defined in Azure Data Factory. It is not created independently, rather, it is dependent on FlowEHR in the implementation details such as names of linked services, paths and so on.
See example in the repo:
{
"name": "PatientsPipeline",
"properties": {
"activities": [
{
"name": "DatabricksPythonActivity",
"type": "DatabricksSparkPython",
"typeProperties": {
"pythonFile": "dbfs:/pipelines/helloworld/artifacts/entrypoint.py",
"libraries": [
{
// Will be uploaded to DBFS by a script in FlowEHR repo
// More libraries can be added here, e.g. from PyPi
"whl": "dbfs:/pipelines/helloworld/artifacts/hello_world-0.0.1-py3-none-any.whl"
}
]
},
"linkedServiceName": {
"referenceName": "ADBLinkedServiceViaMSI", // Needs to be in sync with what is set in FlowEHR
"type": "LinkedServiceReference"
}
}
],
"parameters": {}
},
"type": "Microsoft.DataFactory/factories/pipelines"
}
In there:
linkedServiceName.referenceName
must be set to the same value as it is set on FlowEHR repo.pythonFile
andlibraries
must have a format like this:dbfs:/pipelines/{pipeline_name}/artifacts/{artifact_name}
.
The rest can be copied over from the example above.
- A
Makefile
at the root of the repository with anartifacts
target that will build any artifacts that the pipeline needs to run and put all files that need uploading intoartifacts
directory at the root of the pipeline.
See example in the repo:
artifacts: build
mkdir -p artifacts
cp entrypoint.py artifacts
cp dist/*.whl artifacts/
For a PySpark pipeline, it is recommended to put any library or shared code into a .whl file, as this way it is more straightforward to write automated tests for. There is an entrypoint.py that is required by Azure Data Factory to be present that is just executing the code that is present in the library.
Note that the name of the artifact must match exactly the name specified in pipeline.json
.
To deploy the pipeline in the context of FlowEHR, please follow these steps:
-
Clone FlowEHR repository.
-
Follow the steps outlined in the README to set up your dev environment and deploy FlowEHR. As part of this, you will need to run
make infrastructure-transform
. -
Use one of the following options to reference your pipeline code in FlowEHR repo:
- Copy over your pipeline code (together with a Makefile and an
pipeline.json
) into/transform/pipelines
into your checked-out copy of FlowEHR. (Note that/transform/pipelines
is gitignored in FlowEHR). You should have a directory tree similar to the following:
└── transform └── pipelines ├── my-pipeline │ ├── Makefile │ ├── pipeline.json │ ├── src │ │ └── ... │ ├── tests │ │ └── ... │ └── ...
- In
config.transform.yaml
at the root of the FlowEHR repository, provide a link to your Data Pipeline repository. Then, rebuild your Devcontainer (repositories are checked out on Devcontainer startup).
--- spark_version: 3.3.1 repositories: - git@github.com:tanya-borisova/my-pipeline.git
- Copy over your pipeline code (together with a Makefile and an
-
Re-run
make infrastructure-transform
from the root of FlowEHR while running in your Devcontainer. In the log output, you should see new resources of typeazurerm_data_factory_pipeline.pipeline
being created. If at this stage you don't see the resources being created, double-check that you have bothMakefile
andpipeline.json
at the root of your pipeline directory intransform/pipelines
. -
In Azure Portal, navigate to the instance of Azure Data Factory deployed in your environment. Launch Studio, then on the left under Author find Pipelines. If everything is correct, you should see your pipelines created there. You can click on your pipeline and press Debug to trigger a debug run of the pipeline. If at this stage you don't see your pipeline being created, most likely your
pipeline.json
configuration isn't correct. Review it and try again. -
If necessary, add more pipelines under
/transform/pipelines
and/or inconfig.transform.yaml
by repeating the above steps.
To deploy your data pipeline as part of the CI deployment of your FlowEHR fork, do the following:
-
Update the link to your repository (or repositories) in
config.transform.yaml
, commit and push this change. -
Create a fine-grained personal access token for your organisation with permissions to check out the data pipeline repositories you've created. Put it into a secret named
ORG_GH_TOKEN
. -
Run the
Deploy Infra-Test
action.