Scaling Out Airflow

Note: To learn more about how to scale Airflow check out our Scaling Airflow to optimize performance guide.

1. Key Points About Scaling Airflow

• Virtually unlimited scaling potential
• Use CeleryExecutor and KubernetesExecutor
• Tune parameters to fit your needs
• Easy to scale to more capacity
• Aggregate logging is important

2. High-Level Steps to Scale Airflow

2. Why Scale Apache Airflow®?

• Because workload outgrows your initial infrastructure
  • More DAGs, more Tasks
  • More intensive individual tasks
• Because your core Airflow components need more durability
• To Prepare for more DAGRuns and compute load
• To benefit from elasticity — save money by scaling as needed

4. Symptoms That Mean You’re Ready to Scale

• Many tasks stuck in Queued or Scheduled state
• Unacceptable latency between tasks
• Missing SLAs
• High resource usage on Scheduler or Webserver
• Out of Memory (OOM) errors on Tasks

Principles of Scaling Systems

5. Basics of Scaling Systems

• Vertical Scaling
  • Increase the size of instance
    • (RAM, CPU, etc.)

• Adds more power to an existing worker

• Gives individual tasks more horsepower

• Gets very expensive, very quickly

• If vertical scaling reaches a threshold, think about delegating work to a dedicated distributed processing engine — e.g., Spark, Dask, Ray

• Horizontal Scaling

  • Increase number of instances

    • (RAM, CPU, etc.)

    

    • Adds more nodes to the cluster
    • Increases maximum number of tasks and DAGRuns that the system can handle
    • Fits Airflow’s orchestration model
    • Celery and Kubernetes executors well designed for horizontal scaling

Scaling Airflow as a Distributed Platform

6. Scaling with CeleryExecutor

https://airflow.apache.org/docs/apache-airflow/stable/executor/celery.html

• Allows for easy horizontal scaling

• Runs workers’ processes that process TaskInstances

• To scale, add a new worker process

  • Can be on a new node or an existing node
  • The connection between workers uses Celery broker and metadatabase
  • $AIRFLOW_HOME looks identical to other worker nodes

  

7. Scaling with KubernetesExecutor

https://airflow.apache.org/docs/apache-airflow/stable/executor/kubernetes.html TaskInstances run on K8s pods

• TaskInstance pods are ephemeral
• Each task get its own pod
• No workers’ processes

Parameter Tuning when Scaling Airflow

https://airflow.apache.org/docs/apache-airflow/stable/configurations-ref.html

• parallelism

• max_active_runs_per_dag

• max_active_tasks_per_dag

• worker_concurrency

• Pool size

These parameters control the number of tasks that can be run at a time.

8. Sizing Pools

• Pools are another way that we can limit the amounts of tasks that can run
• Can be used to circumvent executor slots being hogged by heavy DAGs with a lot of tasks
• Limits the number of running and queued tasks (active tasks that are under the control of the executor)
• Groups tasks together to limit the active instances by group

High Availability Airflow Components

Other Airflow Components can Scale!

Scheduler

• Airflow allows for multiple schedulers
• Increases the number of tasks that can be scheduled
• Allows the scheduling platform more stability

Web Server

• Multiple web servers increase the load and capacity of the Web UI

Logging on Distributed Airflow

9. Traits of a Great Logging System

• Aggregated
• Historical
• Indexed
• Searchable

10. Importance of Good Logging Practices

• 85% of problems have clues for a solution in the scheduler logs
• Important, if there are multiple schedulers, to be able to collect all of their logs in one place, since those components are working together on the same solution
• Need to keep a history of the logs in a searchable format in order to diagnose problems and work out solutions
• The first step for debugging is to correlate timestamps between problems in task logs with scheduler logs at the same time

11. Debugging Distributed Airflow

• 90% of problems have clues for a solution in the scheduler logs
• 8% of problems are resource consumption issues
  • Out of memory
  • CPU limited cycles for tasks
• The other 2% is the hard part

How to Scale your Deployment - Demo