Amazon MWAA 24 runs the scheduler, triggerer, API server and Celery workers as Fargate 24 containers and the metadata database on Aurora PostgreSQL 1,289 ; you upload DAGs, plugins and a requirements.txt to an S3 bucket. Each environment pins an image: the newest is Airflow 129 3.3.1 (Python 3.12), offered since 1 September 2026, three weeks after Apache released it. Upgrades stay within a major version, so moving from 2.x to 3.x means a new environment. The environment class sets the size, from mw1.micro (one 1 vCPU, 3 GB container, 3 concurrent tasks, no autoscaling) to mw1.2xlarge (80 tasks); workers autoscale within it.
You pay for every hour the environment exists. In AWS 24 's US East example, a small environment costs $0.49 an hour ($364.56 for a 744-hour month), extra workers $0.055 an hour, and the database $0.10 per GB-month. MWAA Serverless (November 2025) has no environment: workflows are YAML files in the DAG Factory format, each task runs on its own compute under the workflow's IAM role, and you pay for task time (AWS's example: $0.080 an hour, minimum one minute per task). It supports AWS operators only, has no Airflow UI, and suits pipelines of Glue, EMR and Redshift 24 steps, not BookNest's PostgreSQL-and-dbt DAG.