While the single-node Arroyo cluster is useful for testing and development, it is not suitable for production. This page describes how to run a production-ready distributed Arroyo cluster using Arroyo’s built-in scheduler or Kubernetes.

Before attempting to run a cluster, you should familiarize yourself with the Arroyo architecture. We are also happy to support users rolling out their own clusters, so please reach out to us at support@arroyo.systems or on Discord with any questions.

​
Common Setup

​
Postgres

Arroyo relies on a postgres database to store configuration data and metadata. You will need to create a database (by default called arroyo, but this can be configured).

​
Storage

You will need a place to store pipeline artifacts (binaries) and checkpoint data. This needs to be accessible to all nodes in your cluster, including Arroyo services (controller, compiler service) and pipeline workers. Arroyo supports several storage backends, including S3, GCS, and local filesystem. For local testing, a filesystem that’s3 mounted on all nodes is sufficient, but for production you will likely want to use an object store like S3 or GCS. We also support S3-compatible object stores like MinIO and Localstack; endpoints can be set via the s3:: prefix or the AWS_ENDPOINT_URL environment variable.

The storage backend is configured via two environment variables:

  • CHECKPOINT_URL configures where checkpoints are written; for high-availability this should be an object store, but may be a local directory for testing and developing
  • ARTIFACT_URL controls where pipeline artifacts (like compiled UDF libs) are stored

The values for these variables are URLs that specify the storage location. We support a number of ways of specifying these, for example:

  • s3://my-bucket/key/path
  • s3::https://my-custom-s3:1234/my-bucket/key/path
  • https://s3.us-east-1.amazonaws.com/my-bucket
  • file:///my/local/filesystem
  • /my/local/filesystem
  • gs://my-gcs-bucket

​
Prometheus

The Arroyo Web UI can show job metrics to help monitor job progress. To enable this, you will need to set up a Prometheus server. See the prometheus documentation for more details.

Async UDFsKubernetes