Draft. This page is reachable and it is not indexed until Vincent approves it.
Stay on Airflow if your DAGs run reliably and your team knows it. Otherwise, consider Dagster for asset-centric pipelines, Prefect for dynamic Python, Kestra for YAML workflows, Mage for visual block building, and Windmill for scripts plus internal apps. We ran one workflow on Airflow, Dagster, and Prefect; the other three are covered from docs only, as of November 2026.
By Vincent Charles. Last updated . Pricing checked Nov 19, 2026.
Vincent Charles is a data engineer and founder of dataengineers.club and Unchain Data (unchaindata.xyz). He has worked on data at Dailymotion's big data team, Carrefour and Binance, starting in SQL and GIS location intelligence.
Disclosure: dataengineers.club does not sell an orchestrator.
On this page
Key takeaways
- Airflow is current, not legacy. Apache Airflow 3.4.0 was released on PyPI on 2026-11-06 (PyPI).
- The alternatives differ in their core unit. Assets (Dagster), Python flows (Prefect), YAML flows (Kestra), blocks (Mage), and scripts and flows (Windmill).
- Dagster and Prefect share an owner. Prefect announced it was acquiring Dagster Labs on July 13, 2026. Both products continue under their own names (Dagster).
- Licenses differ. Airflow, Dagster, Prefect, Kestra, and Mage are Apache 2.0. Windmill's open-source build is AGPLv3, with enterprise features under a proprietary license (Windmill repo).
- Pricing units differ. Managed offers bill by deployment hours, credits, seats, task runs, execution minutes, or compute units. Compare against your own workload.
- Same workflow, similar code. Airflow, Dagster, and Prefect each needed 92 to 94 lines of pipeline code and each retried as documented. We publish no speed winner.
Why people look for Airflow alternatives
These reasons come from Airflow's own docs and the vendors' published models. We do not cite forum threads.
- Self-hosting has many moving parts. A minimal Airflow installation needs a scheduler, a DAG processor, a DAG bundle, an API server, and a metadata database, usually PostgreSQL or MySQL (Airflow architecture). Someone has to run and upgrade all of it.
- Task-centric vs data-centric thinking. Airflow schedules tasks in DAGs. Airflow 3 added asset-aware scheduling (Airflow assets), but teams that want assets as the core object often look at Dagster.
- Python vs YAML. Airflow DAGs are mostly written in Python (Airflow 3.4 docs also include a page on writing DAGs in YAML). Teams that want declarative YAML across languages look at Kestra.
- Managed cost. Managed Airflow is billed by environment or deployment hours plus workers on Astro, MWAA, and Cloud Composer (Astronomer pricing, MWAA pricing, Composer pricing). Always-on deployments cost money even when idle.
- Azure changes. Azure Data Factory's Workflow Orchestration Manager was retired, and Microsoft points users to Apache Airflow jobs in Microsoft Fabric (Microsoft Learn). A forced move is a natural time to re-evaluate.
Methodology
- Status: test run on 2026-11-19, results from our logs.
- Date of docs and pricing review: 2026-11-19.
- Shared workflow: download the NYTimes COVID-19 US state-level CSV (61,942 rows), clean it, and write a 1,158-row daily summary to DuckDB, with the ingest step failing once on an injected error and recovering through the tool's retry setting.
- Tested: apache-airflow 3.4.0 (SQLite metadata DB,
LocalExecutor,airflow dags test), dagster 1.13.30, and prefect 3.9.4. The Dagster and Prefect runs are the same ones from our Dagster vs Prefect test. - Not tested, docs only: Kestra and Windmill (their self-host paths need Docker or a Java runtime, and neither was installed) and Mage (mage-ai 0.9.81 failed to install on Python 3.13.5 because a dependency needed a Rust toolchain).
- Environment: Debian 13 (Linux 6.12, x86_64), 8 vCPU, 15 GiB RAM, Python 3.13.5. Windows was not tested.
- Times: wall clocks by a tester who already knew the plan; treat them as lower bounds.
- Also not run: the Airflow UI, managed Airflow (MWAA, Cloud Composer, Astro, Fabric), throughput or latency benchmarks, a 1-million-row job, and a port of an existing DAG.
- Repo: repo coming soon.
Comparison matrix
| Tool | Model | Deployment | Open source | Managed offer | Our test (2026-11-19) | Best when |
|---|---|---|---|---|---|---|
| Apache Airflow (incumbent) | Python DAGs of tasks; asset-aware scheduling in Airflow 3 | Self-host; Astro, MWAA, Cloud Composer, Fabric Airflow jobs | Yes, Apache 2.0 | Several vendors | Ran: retried as documented, 92 pipeline lines | You have working DAGs and Airflow skills |
| Dagster | Software-defined assets | Self-host OSS; Dagster+ Serverless or Hybrid | Yes, Apache 2.0 | Dagster+ | Ran: retried as documented, 94 pipeline lines | Tables, dbt, partitions, lineage |
| Prefect | Python flows and tasks | Self-host server; Prefect Cloud with work pools | Yes, Apache 2.0 | Prefect Cloud | Ran: retried as documented, 93 pipeline lines | Dynamic Python, ML, APIs |
| Kestra | Declarative YAML flows, plugins for many languages | Self-host Core; Kestra Cloud; Enterprise self-hosted | Yes, Apache 2.0 | Kestra Cloud | Not tested, docs only | Language-agnostic, YAML-first teams |
| Mage | Pipelines built from Python, SQL, R, and dbt blocks | Self-host OSS; Mage Pro managed | Yes, Apache 2.0 | Mage Pro | Not tested, docs only | Visual, block-by-block building |
| Windmill | Scripts in many languages, composed into flows and apps | Self-host (Docker, Kubernetes); Windmill Cloud | AGPLv3 core, proprietary EE features | Windmill Cloud and EE | Not tested, docs only | Scripts plus internal tools in one place |
Sources: Airflow, Dagster+, Prefect work pools, Kestra pricing, Mage pricing, Windmill repo, and each repo's LICENSE file, as of 2026-11-19.
Pricing as of 2026-11-19
Pricing checked Nov 19, 2026. Every open-source option is free to license. You still pay for servers, a database, and the time to run them.
| Tool | Public starting price or "see source" | Billing unit | What we verified | Source and date |
|---|---|---|---|---|
| Airflow on Astro | Developer deployments from $0.35/hr; Team from $0.42/hr; workers from $0.13/hr; dedicated clusters from $2.40/hr | Deployment hours plus worker hours | Listed rates | Astronomer, 2026-11-19 |
| Airflow on MWAA | See source | Hourly environment by size, plus extra workers, schedulers, web servers, and metadata storage per GB-month | Billing structure only. Not verified: hourly rates did not render in our fetch | AWS, 2026-11-19 |
| Airflow on Cloud Composer 3 | $0.06 per 1,000 milli DCU-hours (that is, $0.06 per DCU-hour) on demand in Iowa (us-central1); $0.054 and $0.048 with 1-year and 3-year committed use discounts | DCU-hours plus database storage | Rates for the default region shown; other regions differ. The page now titles the product Managed Service for Apache Airflow (Gen 3) | Google Cloud, 2026-11-19 |
| Dagster | OSS free. Dagster+ Solo $10/mo plus $0.040 per credit; Starter $100/mo plus $0.035 per credit; Pro custom | Credits (asset materializations plus ops) | Listed rates | Dagster, 2026-11-19 |
| Prefect | OSS free. Cloud Hobby free; Starter $100/mo; Team $100 per user per month; Enterprise custom | Plan fee, per user on Team | Listed rates | Prefect, 2026-11-19 |
| Kestra | Core free. Cloud usage-based; Enterprise annual, contact sales | Task runs plus per-second compute on Kestra's runners (Cloud) | Not on the public page: Cloud rates as of 2026-11-19 | Kestra, 2026-11-19 |
| Mage | OSS free. Starter $29/mo plus $0.0167 per execution minute; Team $100/mo plus $0.0167 per execution minute; Enterprise custom | Plan fee plus execution minutes (1 CPU, 4 GB RAM standard) | Listed rates | Mage, 2026-11-19 |
| Windmill | Community Edition free. Self-hosted Enterprise from $120/mo (calculator: $20 per builder, $10 per non-builder, $50 per standard worker, per month). Self-hosted Pro for individuals and businesses under $250k annual revenue; 40% nonprofit discount. Windmill Cloud tab: Team from $10/mo ($10 per builder, $5 per non-builder, up to 10 seats) and Enterprise from $840/mo (dedicated workers and database), both checked on the rendered page | Seats plus compute units | Calculator defaults; totals scale with seats and workers | Windmill, 2026-11-19 |
Notes: Fabric Apache Airflow jobs are billed on pool uptime in Fabric capacity units: 5 CUs for a Small pool and 10 CUs for a Large pool, plus 0.6 or 1.3 CUs per extra node. The dollar cost depends on your Fabric capacity SKU and region (Microsoft Learn). Annual contracts change most of these numbers.
Pick this if...
| If you need... | Start with | Avoid or postpone |
|---|---|---|
| Stability for DAGs that already work | Stay on Airflow, possibly managed | A rewrite with no measured pain |
| Asset lineage, partitions, dbt-first pipelines | Dagster | Tools without an asset model |
| Wrapping existing Python with retries and tracking | Prefect | YAML-first tools |
| YAML workflows across languages and teams | Kestra | Python-only frameworks |
| Interactive block-by-block pipeline building | Mage | Pure code-in-files tools if your team dislikes them |
| Scripts, flows, and internal apps in one platform | Windmill | Windmill's AGPLv3 build if that license is a problem for your use |
| Durable application workflows, not data DAGs | Temporal (docs; docs only, not tested) | Treating Temporal as a drop-in DAG runner |
The alternatives, one by one
Apache Airflow (the incumbent): stay if...
Airflow is an Apache Software Foundation project for authoring, scheduling, and monitoring DAGs. Version 3.4.0 shipped on 2026-11-06. Airflow 3 has an API server, a DAG processor, asset-aware scheduling, and event-driven scheduling in its docs (Airflow docs). It has the widest choice of managed hosts: Astronomer's Astro, AWS MWAA, Google Cloud Composer, and Apache Airflow jobs in Microsoft Fabric.
- What we ran: Airflow 3.4.0 locally without Docker, using a SQLite metadata database,
LocalExecutor, andairflow dags test. Install took about 1 second on a warm cache (133 packages). The run took 7.6 seconds wall clock, with a 4.7-second DAG run: ingest try 1 failed on the injected error and went toup_for_retry, try 2 succeeded, then transform and write took 0.16 and 0.09 seconds. Outputs matched Dagster and Prefect exactly. Friction: the official 3.4.0 constraints pin duckdb 2.0.0, so this pipeline used a newer DuckDB than the other two;airflow dags listwas empty until we added--local; and Airflow warned that graphviz was missing. We did not start the Airflow UI. - Pros (documented): Apache 2.0; many managed options; large operator and provider ecosystem.
- Cons (documented): many components to self-host; scheduling is task-first.
- Stay if: DAGs run reliably, on-call load is fine, and a managed host already removes the infrastructure work.
Dagster
Dagster centers pipelines on assets, defined as "an object in persistent storage, such as a table, file, or persisted machine learning model" (Dagster docs). You scaffold with create-dagster and run the local UI with dg dev. Partitions, backfills, and asset lineage are built in. Dagster+ comes as Serverless or Hybrid.
Ownership note: Prefect announced on July 13, 2026 that it was acquiring Dagster Labs. Dagster keeps its name, open-source license, and pricing according to both companies (Dagster).
- What we ran: dagster 1.13.30, three assets with a retry policy. The first successful launch took 8.4 seconds wall clock, and ingest retried as configured. From install to first success took 4 minutes 31 seconds, including one failed launch caused by DuckDB rejecting
?placeholders. - Pros (documented): asset model, lineage, partitions; clear local dev loop.
- Cons (documented): the asset mindset is extra structure for side-effect-only jobs; Dagster+ HIPAA not included on Solo or Starter.
- When not to use: mostly dynamic, branch-at-runtime Python with no clear output tables.
Deeper comparison: Dagster vs Prefect.
Prefect
Prefect turns Python functions into flows and tasks with decorators. A flow still runs as a normal script, and Prefect tracks its states, retries, and timeouts (Prefect flows). Deployments add schedules and remote triggering. Work pools choose the infrastructure per deployment: hybrid workers, push to serverless, or Prefect-managed (work pools).
- What we ran: prefect 3.9.4, one flow with three tasks. The first script run took 11.5 seconds wall clock (1.3 seconds of flow time); a run against a local server took 2.6 seconds (1.4 seconds of flow time). Ingest retried in both runs.
- Pros (documented): minimal rewrite of existing Python; task-level retries; flexible infrastructure.
- Cons (documented): fewer built-in data-asset features than Dagster; Cloud run retention is 7 days on Hobby and Starter (pricing).
- When not to use: if partitioned backfills and a data catalog are your main need.
Kestra
Kestra declares workflows in YAML and runs tasks through plugins, including scripts in several languages. The pricing page says Core is free with no limits on flows or executions and lists 2,600+ plugins, a vendor-stated count we did not verify (Kestra pricing). Retries are configured per task in YAML (Kestra retries). The quickstart runs Kestra in one Docker container (quickstart).
- What we ran: not tested, docs only. Kestra needs Docker or a Java runtime, and neither was installed.
- Pros (documented): Apache 2.0; YAML plus code and no-code editors; language-agnostic.
- Cons (documented): SSO, RBAC, secrets manager, and audit logs are not in Core; Cloud rates are not public.
- When not to use: if your team wants pipelines as pure Python code.
Mage
Mage builds pipelines from blocks of Python, SQL, R, or dbt in an interactive editor, with schedules, sensors, and backfills listed on its pricing page (Mage pricing). The open-source package is mage-ai on PyPI (0.9.81, released 2026-11-18), and the repo is Apache 2.0. Mage's docs now lead with Mage Pro, the managed product (Mage docs).
- What we ran: not tested, docs only. mage-ai 0.9.81 failed to install on Python 3.13.5: pydantic-core 2.18.4 needed a Rust build and no toolchain was set up.
- Pros (documented): interactive building; data integration and orchestration in one tool; block retries (docs).
- Cons (documented): multi-environment workspaces, CI/CD deployments, cluster monitoring, granular RBAC, audit logs, and managed autoscaling are Mage Pro features, not OSS (Mage OSS vs Pro).
- When not to use: if you want pipelines reviewed as plain code files with no editor in the loop.
Windmill
Windmill runs scripts in TypeScript, Python, Go, Bash, SQL, and other languages, and composes them into flows and apps (scripts quickstart). Flow steps can retry with static or exponential delays (Windmill retries). Self-hosting uses Docker Compose or Kubernetes (self-host).
- What we ran: not tested, docs only. Windmill self-hosting needs Docker or Kubernetes, and Docker was not installed.
- Pros (documented): many languages; flows and internal apps together; unlimited executions on the free edition.
- Cons (documented): AGPLv3 for the open-source build; the README says serving it as a managed service or embedding it in your product needs a commercial license (repo).
- When not to use: if your legal team does not accept AGPLv3 and you will not buy a license.
Windmill also publishes its own "Airflow alternatives" post. It is a vendor page that includes Windmill; we did not use its claims.
Migration notes: moving off Airflow
These steps come from the tools' documented models. We do not estimate effort in days.
- Inventory your DAGs. List each DAG's schedule, operators, connections, variables, sensors, and SLAs. Operators that call outside systems need an equivalent in the new tool.
- Map concepts, not code. An Airflow task maps roughly to a Prefect task, a Dagster asset or op, a Kestra task, a Mage block, or a Windmill script. Asset-centric tools may merge several tasks into one asset.
- Move secrets deliberately. Airflow connections and variables do not transfer automatically. Recreate them in the target's secrets system.
- Run both in parallel. Point the new tool at a separate schema or bucket, compare outputs, then switch.
- Disable the Airflow DAG last, after the new job has run on schedule for a period you choose.
- Consider a managed Airflow host first if the only pain is ops. Moving hosts keeps your DAG code. Azure users on Workflow Orchestration Manager already have a documented path to Fabric (Microsoft Learn).
- Rewrite tests. DAG-level tests do not carry over. Write tests in the target's style before cutover.
We did not port an existing Airflow DAG; we wrote the workflow fresh in each tool. Line counts were close: 92 for Airflow, 94 for Dagster, and 93 for Prefect, counting pipeline lines only. Pinned dependencies differed: Airflow's official constraints forced duckdb 2.0.0, while the Dagster and Prefect runs used 1.3.2, so check every shared library against the target tool's constraints.
FAQ
What is better than Airflow?
Nothing is better for every team. As of 2026-11-19, Dagster suits asset-centric data pipelines, Prefect suits dynamic Python, Kestra suits YAML workflows, Mage suits interactive building, and Windmill suits scripts plus internal apps. If Airflow already works for you, staying, or moving to a managed host, may cost less than a rewrite. See the decision table above.
Can dbt replace Airflow?
No. dbt is a transformation tool: it builds, tests, and documents SQL models in your warehouse ([dbt docs](https://docs.getdbt.com/docs/introduction)). It does not orchestrate ingestion, API calls, or ML jobs. Many teams run dbt from an orchestrator. dbt Labs merged with Fivetran on June 1, 2026 ([Fivetran](https://www.fivetran.com/blog/fivetran-recognized-as-a-leader-in-the-2026-gartner-r-magic-quadrant-tm-for-data-integration-tools)); that does not make dbt an orchestrator.
Which is better, NiFi or Airflow?
They do different jobs. Apache NiFi is a flow-based system "to process and distribute data," with data provenance tracking ([NiFi](https://nifi.apache.org/)). Airflow schedules and monitors batch DAGs of tasks. NiFi fits continuous data routing between systems; Airflow fits scheduled pipelines. We did not test NiFi, so we name no winner.
Is Airflow still used?
Yes. As of 2026-11-19, Apache Airflow 3.4.0 was released on 2026-11-06 ([PyPI](https://pypi.org/project/apache-airflow/)), and managed Airflow is sold by Astronomer, AWS (MWAA), Google Cloud (Cloud Composer), and Microsoft (Fabric Apache Airflow jobs). We do not quote market-share figures, because we have no independent measurement.
What are the open-source Airflow alternatives?
Dagster, Prefect, Kestra, and Mage are Apache 2.0 licensed, checked on 2026-11-19 against their repos and PyPI. Windmill's open-source build is AGPLv3, with enterprise features under a proprietary license ([Windmill repo](https://github.com/windmill-labs/windmill)). All of them can be self-hosted. "Free" means free to license, not free to operate.
What are the Airflow options on AWS, GCP, and Azure?
As of 2026-11-19: on AWS, Amazon MWAA and Astro through the AWS Marketplace; on GCP, Cloud Composer; on Azure, Apache Airflow jobs in Microsoft Fabric, which replaced Azure Data Factory's Workflow Orchestration Manager ([Microsoft Learn](https://learn.microsoft.com/en-us/fabric/data-factory/apache-airflow-jobs-concepts)). Fabric's page lists Airflow 2.10.5 support, not Airflow 3.
Is this comparison current for 2026?
The docs and pricing review is dated 2026-11-19, and versions are pinned in the methodology box. The shared workflow results are pending and will carry their own run date. Recheck pricing pages before you buy, since all of them change.
Sources
All links checked on 2026-11-19.
- Apache Airflow architecture overview: https://airflow.apache.org/docs/apache-airflow/stable/core-concepts/overview.html
- Apache Airflow asset definitions: https://airflow.apache.org/docs/apache-airflow/stable/authoring-and-scheduling/assets.html
- PyPI, apache-airflow: https://pypi.org/project/apache-airflow/
- Astronomer pricing: https://www.astronomer.io/pricing/
- Amazon MWAA pricing: https://aws.amazon.com/managed-workflows-for-apache-airflow/pricing/
- Google Cloud Composer pricing: https://cloud.google.com/composer/pricing
- Microsoft Fabric Apache Airflow jobs: https://learn.microsoft.com/en-us/fabric/data-factory/apache-airflow-jobs-concepts
- Dagster pricing: https://dagster.io/pricing
- Dagster assets: https://docs.dagster.io/guides/build/assets
- Dagster is now part of Prefect: https://dagster.io/prefect
- Prefect pricing: https://www.prefect.io/pricing
- Prefect flows: https://docs.prefect.io/v3/concepts/flows
- Prefect work pools: https://docs.prefect.io/v3/concepts/work-pools
- Kestra pricing: https://kestra.io/pricing
- Kestra retries: https://kestra.io/docs/workflow-components/retries
- Kestra quickstart: https://kestra.io/docs/getting-started/quickstart
- Mage pricing: https://www.mage.ai/pricing
- Mage docs: https://docs.mage.ai/getting-started/setup
- Mage block retries: https://docs.mage.ai/orchestration/pipeline-runs/retrying-block-runs
- Mage OSS vs Pro: https://docs.mage.ai/migrations/mage-oss-to-mage-pro
- Microsoft Fabric Apache Airflow job pricing: https://learn.microsoft.com/en-us/fabric/data-factory/pricing-apache-airflow-job
- Windmill pricing: https://www.windmill.dev/pricing
- Windmill retries: https://www.windmill.dev/docs/flows/retries
- Windmill scripts quickstart: https://www.windmill.dev/docs/getting_started/scripts_quickstart
- Windmill self-hosting: https://www.windmill.dev/docs/advanced/self_host
- Windmill repo and license: https://github.com/windmill-labs/windmill
- dbt introduction: https://docs.getdbt.com/docs/introduction
- Apache NiFi: https://nifi.apache.org/
- Temporal workflows docs: https://docs.temporal.io/workflows
- Fivetran blog (merger completion date): https://www.fivetran.com/blog/fivetran-recognized-as-a-leader-in-the-2026-gartner-r-magic-quadrant-tm-for-data-integration-tools
Get the newsletter
One email when a comparison on this page is actually finished.