Configure resources for Airflow components on Astronomer Software Deployments
Use this document to configure resource usage for a Deployment's executor , webserver, scheduler and triggerer components.
Select an executor
The Airflow executor works closely with the Airflow scheduler to determine what resources complete tasks as they queue. The main difference between executors is their available resources and how they utilize those resources to distribute work.
Astronomer supports three executors:
Though it largely depends on your use case, Astronomer recommends the Local executor for development environments and the Celery or Kubernetes executors for production environments operating at scale.
For a detailed description of each executor, see Airflow executors explained.
Select a resource strategy
A Deployment's Resource Strategy defines how you can allocate CPU and memory to the Deployment's Airflow components. Astronomer Software offers two different resource strategies: Custom Resources and Astronomer Units (AUs).
An AU is equivalent to 0.1 CPU and 0.375 GiB of memory. If you set your resource strategy to Astronomer Units, you can only scale components based on this resource ratio. Components must use the same AU value for both CPU and memory.
If you set your resource strategy to Custom Resources, you can freely set CPU and memory for each component without a predetermined ratio. See Customize resource usage.
If you still want a constant ratio of CPU to memory, but also want to change the specific ratio, you can change the amount of resources an AU represents. See Overprovision Deployments.
Scale core resources
Apache Airflow requires four primary components:
- The Webserver
- The Scheduler
- The executor (and the workers it runs)
- The triggerer
To scale these resources, adjust the corresponding slider in the Software UI to increase its available resources. The units associated with these sliders will vary based on your resource strategy.
Read the following sections to help you determine which core resources to scale and when.
Airflow webserver
The Airflow webserver is responsible for rendering the Airflow UI, where users can monitor DAGs, view task logs, and set various non-code configurations.
If a function within the Airflow UI is slow or unavailable, Astronomer recommends increasing the resources allocated towards the webserver.
Scheduler
The Airflow scheduler is responsible for monitoring task execution and triggering downstream tasks once dependencies have been met.
If you experience delays in task execution, which you can track via the Gantt Chart view of the Airflow UI, Astronomer recommends increasing the resources allocated towards the scheduler.
To set alerts that notify you via email when your Airflow scheduler is underprovisioned, configure an Airflow alert.
Scheduler count
Airflow 2.0 comes with the ability for users to run multiple schedulers concurrently to ensure high-availability, zero recovery time, and faster performance. You can provision up to 4 schedulers on any Deployment.
Each individual scheduler will be provisioned with the resources specified in Scheduler Resources. For example, if you set the CPU figure in Scheduler Resources to 5 CPUs and set Scheduler Count to 2, your Airflow Deployment will run with 2 Airflow schedulers using 5 CPUs each for a total of 10 CPUs.
To increase the speed at which tasks are scheduled and ensure high-availability, Astronomer recommends provisioning 2 or more Airflow schedulers for production environments.
Triggerer
Airflow 2.2 introduces the triggerer, which is a component for running tasks with deferrable operators. Like the scheduler, the triggerer is highly-available: If a triggerer shuts down unexpectedly, the tasks it was deferring can be recovered and moved to another triggerer.
By adjusting the Triggerer slider in the Software UI, you can provision up to 2 triggerers on any Deployment running Airflow 2.2+. To take advantage of the Triggerer's high availability, we recommend provisioning 2 triggerers for production Deployments.
(Kubernetes executor only) Set extra capacity
On Astronomer, resources required for the KubernetesPodOperator or the Kubernetes Executor are set as Extra Capacity.
The Kubernetes executor and KubernetesPodOperator each spin up an individual Kubernetes pod for each task that needs to be executed, then spin down the pod after that task is completed.
The amount of CPU and Memory allocated to Extra Capacity maps to resource quotas on the Kubernetes Namespace in which your Airflow Deployment lives on Astronomer. More specifically, Extra Capacity represents the maximum possible resources that could be provisioned to a pod at any given time.
Resources allocated to Extra Capacity do not affect scheduler or webserver performance and do not represent actual usage.
(Celery executor only) Configure workers
To optimize for flexibility and availability, the Celery executor works with a set of independent Celery workers across that it can delegate tasks. On Astronomer, you can configure your Celery workers to fit your use case.
Worker count
By adjusting the Worker Count slider, users can provision up to 20 Celery workers on any Airflow Deployment.
Each individual worker will be provisioned with the resources specified in Worker Resources. If you set the CPU figure in Worker Resources to 5 CPUs and set Worker Count to 3, for example, your Airflow Deployment will run with 3 Celery workers using 5 CPUs each for a total of 15 CPUs.
Worker termination grace period
On Astronomer, Celery workers restart after every code deploy to your Airflow Deployment. This makes sure that workers execute with the most up-to-date code. To minimize disruption during task execution, however, Astronomer supports the ability to set a Worker Termination Grace Period.
If a deploy is triggered while a Celery worker is executing a task and Worker Termination Grace Period is set, the worker will continue to process that task up to a certain number of minutes before restarting itself. By default, the grace period is ten minutes.