Mastering Apache Airflow: My Essential Best Practices for Robust Data Orchestration

Harry @DataSmiles
5 min readNov 14, 2023

Apache Airflow stands as a formidable tool in the realm of data engineering, often serving as the linchpin for data workflows in many teams — a role it fulfills in ours as well. Mastering Airflow, however, is no small feat. While Airflow’s own documentation offers a solid foundation of best practices, my experience as a hands-on practitioner provides an additional layer of insight, potentially more aligned with the practical challenges you might face.

In this blog, I aim to share the wealth of knowledge and best practices we’ve accumulated over years of using Airflow as our primary orchestration tool.

Our journey with Airflow has been transformative. When I first assumed responsibility for our Airflow infrastructure, it was a single-server setup with a diverse array of DAGs and little standardization. Fast forward to today, we’ve transitioned to an ECS-based system, rewritten our DAGs for consistency, and embraced Kubernetes Executors — though not the K8sOperator. Our most recent milestone has been deploying Airflow across two additional data centers, multiplying our implementation’s complexity.

Key Best Practices from Our Experience

  1. Implementing a standardized data pipeline

Implementing a standardized data pipeline has significantly streamlined our processes. But what exactly constitutes a “standard data pipeline”? In essence, it’s a framework that allows for the uniform handling of data from diverse sources and in various formats. This standardization enables us to reuse a substantial amount of code, simplifying the integration of new data sources and operations, and reducing the likelihood of errors.

Our standardized pipeline process typically involves the following steps:

  • Data Retrieval: We start by fetching data through HTTP requests from various APIs.
  • Data Parsing: The retrieved JSON responses are then parsed and transformed into internally defined data objects.
  • Data Storage: These data objects are subsequently dumped into local Parquet files, an efficient file format for large datasets.
  • Data Upload: Next, we upload these Parquet files to a specific S3 path allocated for each DAG and task run, ensuring organized and accessible data storage.
  • Data Integration in Snowflake: In Snowflake, we check for the existence of the target table. If it doesn’t exist, we create a new table based on the data class defined in the data parsing step.
  • Data Transfer: The data from the S3 files is then copied to a staging table within Snowflake.
  • Data Merging: From the staging table, we merge the data into the main target table, ensuring up-to-date and comprehensive datasets.
  • Data Archiving: Finally, we archive the S3 files, which are scheduled for automatic deletion after 30 days to maintain storage efficiency.

For incorporating a new data source, the process is streamlined: we only need to define the data object and develop the code for data retrieval from the new API. This standardized approach is also adaptable for loading data from other databases or Kafka streams, highlighting its versatility and efficiency.

2. Use a Distributed Architecture with Containers:

Adopting a containerized, distributed architecture in Airflow, such as through Docker and Kubernetes, enhances system stability and scalability. This approach isolates different components of the Airflow setup, like the scheduler and workers, preventing a single point of failure and facilitating easier scaling and updates. It also allows for better resource management and deployment flexibility, adapting to varying workloads efficiently.

3. Make Tasks Idempotent and Resilient:

Designing idempotent tasks ensures that even if a task is executed multiple times, it won’t produce duplicate results, maintaining data integrity. Resilience in tasks is crucial for handling interruptions gracefully, ensuring they can resume or restart without data loss or corruption. This feature is particularly important in distributed environments where network issues or hardware failures are common. Building these characteristics into your tasks results in a more reliable and robust data pipeline.

4. Implement Your Own Operators:

Creating custom Operators in Airflow for specific ETL tasks allows for more tailored data processing, especially when interacting with various external systems. These Operators can encapsulate complex logic, API calls, and error handling specific to each external service, thereby simplifying your DAGs and making them more readable and maintainable.

5. Periodically Clean Up Airflow’s Metadata Database:

Regular maintenance of the Airflow metadata database, such as purging old task logs and historical data, can significantly improve the performance of the Airflow webserver and scheduler. This practice prevents the database from becoming a bottleneck due to excessive size and helps maintain quick response times in the UI and scheduler operations.

6. Managing Resources Wisely:

Efficient resource management in Airflow involves strategic practices like storing all connection information in Airflow Connections, avoiding the use of Variables for frequent access, and minimizing resource-heavy operations within DAG files. This approach reduces the load on the database and the frequency of expensive operations, leading to better overall performance of your Airflow environment.

7. Separate Airflow Infrastructure from DAG Deployments:

Decoupling the deployment of Airflow’s infrastructure from the DAGs is a significant step towards a more maintainable and scalable system. This separation allows for independent updates and changes to either the DAGs or the Airflow components without impacting the other. Achieving this separation often requires careful planning and a robust CI/CD pipeline to ensure seamless updates and deployments.

8. Have Task Timeout Setup in Airflow:

Implementing task timeouts is crucial for managing long-running tasks. By setting a reasonable timeout threshold (typically 2 to 3 times the average task duration), you can prevent tasks from running indefinitely, which could otherwise lead to resource wastage and potential bottlenecks. This practice also helps in quickly identifying and addressing tasks that are stuck or taking unusually long to complete, aiding in maintaining the efficiency and reliability of your workflows.

9. Ensure Runtime Isolation for Prolonged Jobs to Maintain Airflow Infrastructure Stability:

It’s critical to guarantee that the runtime of long-running jobs is isolated to prevent disruptions in Airflow infrastructure, particularly avoiding any impact on worker pods.

  • Utilizing Persistent Volumes for DAG Storage: Implement a Persistent Volume (PV) and a corresponding Persistent Volume Claim (PVC) within Kubernetes for DAG storage. This setup, where the PV is integrated into both the Airflow scheduler and worker pods, ensures seamless accessibility and management of DAGs.
  • Simplified DAG Updates: When modifying DAGs, simply revise the files in this persistent volume. The Airflow scheduler, designed to monitor these changes, will automatically recognize and apply the updates to the DAGs.
  • Incorporating a CI/CD Pipeline for Automated Updates: Establish a CI/CD pipeline dedicated to the Airflow environment in Kubernetes. This system streamlines the update process, automatically synchronizing changes from your version control to the Airflow setup.
  • Automated DAG Refresh through CI/CD Integration: With every update or modification made to a DAG in your version control, the CI/CD pipeline can be configured to automatically transfer these changes into your Kubernetes environment, ensuring your DAGs are always current.”

Certainly, here’s a more concise conclusion for your article:

Final Thoughts: Harnessing the Power of Airflow

As we conclude our deep dive into Apache Airflow best practices, it’s clear that mastering Airflow is an ongoing journey of optimization and learning. The strategies we’ve outlined — from embracing containerized architectures and crafting robust tasks to smart resource management and DAG deployment techniques — are pivotal in building a resilient, efficient, and scalable data orchestration system.

Remember, the true power of Airflow lies in its flexibility and adaptability. By continuously refining your approach and staying abreast of best practices, you ensure that your data pipelines are not just functional but are also future-proofed against the ever-evolving landscape of data engineering.

Let’s keep pushing the boundaries with Airflow, transforming data challenges into opportunities for innovation and success.



Harry @DataSmiles

Chief Solution Architect @ Data Smiles. Former Engineering Manager of Data team at ActiveCampaign.