Apache Airflow Provider for Bacalhau

This is bacalhau-airflow, a Python package that integrates Bacalhau with Apache Airflow. The benefit is twofold. First, thanks to this package you can now write complex pipelines for Bacalhau. For instance, jobs can communicate their output's CIDs to downstream jobs, that can use those as inputs. Second, Apache Airflow provides a solid solution to reliably orchestrate your DAGs.

You may try this out using a local devstack until https://github.com/bacalhau-project/bacalhau/issues/2038 has been fixed. Please set the following environment variables AIRFLOW_VAR_BACALHAU_API_HOST, AIRFLOW_VAR_BACALHAU_API_PORT.

Features

  • Create Airflow tasks that run on Bacalhau (via custom operator!)

  • Support for sharded jobs: output shards can be passed downstream (via XComs)

  • Coming soon...

Requirements

Installation

The integration automatically registers itself for Airflow 2.3+ if it's installed on the Airflow worker's Python.

From pypi

pip install bacalhau-airflow

From source

Clone the public repository:

git clone https://github.com/bacalhau-project/bacalhau/

Once you have a copy of the source, you can install it with:

cd integration/airflow/
pip install .

Worked example

Setup

In a production environment you may want to follow the official Airflow's instructions or pick one of the suggested hosted solutions.

If you're just curious and want to give it a try on your local machine, please follow the steps below.

First, install and initialize Airflow:

$ pip install apache-airflow
export AIRFLOW_HOME=~/airflow
$ airflow db init

Then, we need to point Airflow to the absolute path of the folder where your pipelines live. To do that we edit the dags_folder field in ${AIRFLOW_HOME}/airflow.cfg file. In this example I'm going to use the hello_world.py DAG shipped with this repository; for the sake of completeness, the next section will walk you through the actual code.

My config file looks like what follows:

[core]
dags_folder = /Users/enricorotundo/bacalhau/integration/airflow/example_dags
...

Optionally, to reduce clutter in the Airflow UI, you could disable the loading of the default example DAGs by setting load_examples to False.

Finally, we can launch Airflow locally:

airflow standalone

Example DAG: chaining jobs

While Airflow's pinwheel is warming up in the background, let's take a look at the hello_world.py breakdown below.

In brief, the first task of this DAG prints out "Hello World" to stdout, then automatically pipe its output into the subsequent task as an input file. The second task will simply print out the content of its input file.

All you need to import from this package is the BacalhauSubmitJobOperator. It allows you to submit a job spec comprised of the usual fields such as engine, image, etc.

from datetime import datetime
from airflow import DAG
from bacalhau_airflow.operators import BacalhauSubmitJobOperator

This operator supports chaining multiple jobs without the need to manually pass any CID along, in this regards a special note goes to the input_volumes parameter (see task_2 below). Every time the operator runs a task, it stores a comma-separated string with the output shard-CIDs in an internal key-value store under the cids key. Thus, downstream tasks can read in those CIDs via the input_volumes parameter.

All you need to do is (1) use the XComs syntax (in curly brackets) to specify the "sender" task ids and the cids key (e.g. {{ task_instance.xcom_pull(task_ids='task_1', key='cids') }}), (2) define a target mount point separated by a colon (e.g. :/task_1_output).

Lastly, we define task dependencies simply with task_1 >> task_2. To learn more about Airflow's DAG syntax please check out this page.

with DAG("bacalhau-helloworld-dag", start_date=datetime(2023, 3, 1)) as dag:
    task_1 = BacalhauSubmitJobOperator(
        task_id="task_1",
        api_version="V1beta1",
        job_spec=dict(
            engine="Docker",
            verifier="Noop",
            publisher="IPFS",
            docker=dict(
                image="ubuntu",
                entrypoint=["echo", "Hello World"],
            ),
            deal=dict(concurrency=1, confidence=0, min_bids=0),
        ),
    )

    task_2 = BacalhauSubmitJobOperator(
        task_id="task_2",
        api_version="V1beta1",
        input_volumes=[
            "{{ task_instance.xcom_pull(task_ids='task_1', key='cids') }}:/task_1_output",
        ],
        job_spec=dict(
            engine="Docker",
            verifier="Noop",
            publisher="IPFS",
            docker=dict(
                image="ubuntu",
                entrypoint=["cat", "/task_1_output/stdout"],
            ),
            deal=dict(concurrency=1, confidence=0, min_bids=0),
        ),
    )

    task_1 >> task_2

Run it

Now that we understand what the example DAG is supposed to do, let's just run it! Head over to http://0.0.0.0:8080, where Airflow UI is being served. The screenshot below shows our hello world has been loaded correctly.

When you inspect a DAG, Airflow will render a graph depicting a color-coded topology (see image below). For active (i.e. running) pipelines, this will be useful to oversee the status of each task.

To trigger a DAG please enable the toggle shown below.

When all tasks have been completed, we want to fetch the output of our pipeline. To do so we need to retrieve the job-id of the last task. Click on a green box in the task_2 line and then open the XCom tab.

Here we find the bacalhau_job_id. Select that value and copy it into your clipboard.

Lastly, we can use the bacalhau cli get command to fetch the output data as follows:

$ bacalhau job get 8fdab73b-00fd-4d13-941c-8ba002f8178d
Fetching results of job '8fdab73b-00fd-4d13-941c-8ba002f8178d'...
...
Results for job '8fdab73b-00fd-4d13-941c-8ba002f8178d' have been written to...
/tmp/dag-example/job-8fdab73b

$ cat /tmp/dag-example/job-8fdab73b/combined_results/stdout
Hello World

Development

pip install -r dev-requirements.txt

Unit tests

tox

You can also skip using tox and run pytest on your own dev environment.