Choosing — The Practical Decision Framework

The honest comparison

None of the three is universally better. Pick based on team shape and constraints, not on Twitter discourse. Here's the working table.

	Airflow	Dagster	Prefect
Mental model	Tasks → DAG	Assets → graph emerges	Functions → flows
Deployment cost	High (DB + scheduler + workers + webserver)	Low (single process to start)	Lowest (single process; Cloud option)
Ecosystem of operators	Largest (1,500+)	Medium	Medium
Best for teams that…	Already have ops; need many integrations	Think in data assets and want lineage front-and-center	Want orchestration that feels like a Python library
Async / concurrency story	Fine	Good	Best (native async)
UI maturity	Mature, dense	Modern, lineage-first	Modern, flow-first
Learning curve	Steep	Medium	Gentle

The defaults I actually recommend

Small/new team, starting from zero: Dagster. The asset model maps to how data teams already think, and the deployment story is the gentlest of the three.
Big team with existing infrastructure and many vendor integrations: Airflow. The ecosystem advantage is real.
Engineering team that wants orchestration to feel like a library, not a platform: Prefect. The hybrid execution model is a real advantage for regulated/security-sensitive shops.

And one important non-answer: don't agonize over the choice. All three are good enough that you'll get most of the value from any of them. The discipline of writing pipelines as DAGs/assets/flows is what makes the work better — the specific tool is the smaller decision.

Code

Same pipeline, three frameworks — pick the one that matches your team·text

Airflow:   @dag → @task functions → DAG file in dags/
Dagster:   @asset functions → repository → dagster dev
Prefect:   @flow + @task functions → python flow.py

All three give you: scheduling, retries, UI, run history, alerting, backfills.
Pick based on team shape, not benchmarks. The decision will not haunt you.

Exercise

Pick one of the three frameworks (Dagster is the gentlest place to start). Build a working three-task pipeline that extracts, transforms, and loads any small dataset. Schedule it. Make it fail on purpose, watch the retry happen, then fix it. The goal is not to ship a production pipeline — it's to feel the loop.