← Back to docs
Recipe

AWS Step Functions primer

AWS Step Functions orchestrate distributed services as visual state machines. This primer walks through the moving parts so you can wire Meridian model calls into a production-grade workflow with retries, parallel branches, and observable history.

1.State machine anatomy

A Step Functions state machine is an Amazon States Language (ASL) document. It declares states — Task, Choice, Parallel, Map, Wait, Pass, Succeed, Fail — and transitions. Standard workflows persist execution history for up to one year and bill per state transition; Express workflows bill per duration and are sub-second ideal for high-volume Meridian fanouts.

  • Task states invoke Lambda, ECS, or any AWS SDK action directly.
  • Choice states branch on JSONPath expressions against the running payload.
  • Map states iterate a list with bounded concurrency — perfect for batch inference.

2.Calling Meridian from a Task state

Wrap the Meridian gateway in a thin Lambda or use the HTTPS Task integration. The snippet below defines a single Task state that posts a chat completion request tollm.getnimbus.netand routes the response into the next state.

{
  "Comment": "Meridian completion via Step Functions",
  "StartAt": "CallMeridian",
  "States": {
    "CallMeridian": {
      "Type": "Task",
      "Resource": "arn:aws:states:::http:invoke",
      "Parameters": {
        "ApiEndpoint": "https://llm.getnimbus.net/v1/chat/completions",
        "Method": "POST",
        "Authentication": {
          "ConnectionArn": "arn:aws:events:us-east-1:000:connection/meridian"
        },
        "RequestBody": {
          "model": "azure/model-router",
          "messages.$": "$.messages"
        }
      },
      "Retry": [{
        "ErrorEquals": ["States.ALL"],
        "IntervalSeconds": 2,
        "MaxAttempts": 3,
        "BackoffRate": 2.0
      }],
      "End": true
    }
  }
}

3.Observability and cost guardrails

Enable CloudWatch Logs at the ALL level during development and switch to ERROR-only in production. Use X-Ray tracing on the state machine to see span-level latency for every Meridian call. Set a Map stateMaxConcurrencyto cap concurrent gateway hits, and front the workflow with an SQS queue so bursts shed gracefully instead of failing executions.