Worker Service¶

Status: migrates (ADR-0014). Verified on real AWS (2026-10-07 run).

A Worker Service (worker-svc) reads messages from an SQS queue. Copilot creates an events queue, a dead-letter queue, one queue per topic it subscribes to, and SNS subscriptions from other services' topics. It scales on backlog per task, a metric that a small Lambda publishes every minute.

What is imported¶

CloudFormation	Terraform	Notes
`AWS::SQS::Queue`, `AWS::SQS::QueuePolicy`	`aws_sqs_queue`, `aws_sqs_queue_policy`	`prevent_destroy` on queues
`AWS::SNS::Subscription`	`aws_sns_topic_subscription`	filter policy and raw delivery checked against live
`AWS::KMS::Key`	`aws_kms_key`	the key that encrypts the queues
`BacklogPerTaskCalculatorFunction`	`aws_lambda_function`	imported, not deleted; its code location is ignored
`BacklogPerTaskScheduledRule`	`aws_cloudwatch_event_rule` + `aws_cloudwatch_event_target`	every minute
`PermissionToInvokeBacklogPerTaskCalculatorLambda`	`aws_lambda_permission`
Autoscaling policies on `BacklogPerTask`	`aws_appautoscaling_policy`
Service, task definition, roles, alarms, topics	as for any service

Why the backlog Lambda is kept¶

Every other Lambda in a Copilot stack handles a custom resource and is deleted after the migration. This one is not a custom resource: it computes ceil(messages / running tasks) for each queue and publishes it as the metric the scaling policies track. Deleting it would leave the service running at whatever size it was, with no error anywhere. ecsodus imports it, its role, log group, schedule and permission.

Cleaned up by hand¶

EnvControllerAction and DynamicDesiredCountAction (it only set the starting task count), and their Lambdas.

Watch out for¶

Topics owned by other services. A subscription points at another service's topic. The subscription moves to Terraform; the topic stays wherever its service is.
EFS from another stack. Copilot's worker render can mount a file system that a separate stack exports. ecsodus reads exactly the exports the templates import by name, and nothing else.
Runtime deadline. The backlog Lambda runs on nodejs20.x. AWS blocks updates to that runtime from 2027-08-31; move it to a supported runtime before then (the generated Terraform carries a note).