Ask AI

Running multiple agents#

This guide is applicable to Dagster+.

Each Dagster+ full deployment (e.g., prod) needs to have at least one agent running. A single agent is adequate for many use cases, but you may want to run multiple agents to provide redundancy if a single agent goes down.

Running multiple agents in the same environment#

To run multiple agents in the same environment (e.g., multiple Kubernetes agents in the same namespace), you can set the number of replicas in the configuration for your particular agent type:

In Docker#

In Docker, you can set the number of replicas for a service in the docker-compose.yaml file if the deployment mode is set to replicated (which is the default):

      mode: replicated
      replicas: 2

Running multiple agents in different environments#

To run multiple agents in an environment where each agent can not access the others' resources (for example, multiple Kubernetes namespaces or different clusters), enable the isolated_agents option. This is supported for all agent types.

In Docker#

Add the following to the dagster.yaml file:

  enabled: true

  # <your other config>
  agent_label: "My agent" # optional

Routing requests to specific agents#

Agent queues are a Dagster+ Pro feature.

Every Dagster+ agent serves requests from one or more queues. By default, requests for each code location are placed on a default queue and your agent will read requests only from that default queue.

In some cases, you might want to route requests for certain code locations to specific agents. For example, routing requests for one code location to an agent running in an on-premise data center, but then routing requests for all other code locations to an agent running in AWS.

To route requests for a code location to a specific agent, annotate the code locations with the name of a custom queue and configure an agent to serve only requests for that queue.

Step 1: Define an agent queue for the code location#

First, set an agent queue for the code location in your dagster_cloud.yaml:

# dagster_cloud.yaml

  - location_name: data-eng-pipeline
      package_name: quickstart_etl
    executable_path: venvs/path/to/dataengineering_spark_team/bin/python
    agent_queue: special-queue

Step 2: Configure an agent to handle the agent queue#

Next, configure an agent to handle your agent queue.

In Docker#

Add the following to your project's dagster.yaml file:

  include_default_queue: True # Continue to handle requests for code locations that aren't annotated with a specific queue
    - special-queue