Apache Airflow vs Fivetran: Build or Buy Pipelines (2026)

Q: Who publishes the Fivetran provider for Airflow?

The maintained package, airflow-provider-fivetran-async, is published in Astronomer's GitHub organisation under Apache-2.0 and was announced by Fivetran as built with Astronomer. It is not part of the Apache Airflow project's own providers. The older airflow-provider-fivetran repository under Fivetran's organisation is archived.

Quick verdict

Short answer

Choose Fivetran when your sources are covered by its connectors and you would rather pay for usage than build and maintain extraction code: it handles schema changes, incremental updates and scheduling for you. Choose Apache Airflow when you need to orchestrate many kinds of work across systems, when sources are custom or unsupported, or when licence cost matters more than engineering time. They are often combined: Fivetran loads raw data, and Airflow triggers or waits for Fivetran syncs before running transformations and checks, using the Fivetran provider published in Astronomer's GitHub organisation.

How we know: This comparison is research-based: features, versions, licences, provider status and pricing were checked against the Apache Airflow documentation, PyPI and GitHub metadata for the Fivetran Airflow providers, Fivetran's documentation, pricing page and press releases in October 2026. We have not run either product, measured performance or estimated anyone's bill.

Apache Airflow is an Apache Software Foundation project, licensed under Apache-2.0, for authoring, scheduling and monitoring workflows written as Python DAGs (directed acyclic graphs of tasks). It orchestrates; the data movement happens in whatever each task calls. The current release is 3.3.2 (17 September 2026), part of the Airflow 3 line that began in April 2025 with DAG versioning, asset-based scheduling, a new UI and a Task SDK. Airflow is free to self-host and is sold as a managed service by Astronomer, Amazon (MWAA) and Google Cloud (Managed Service for Apache Airflow, formerly Cloud Composer), among others.

Fivetran is a commercial, fully managed data movement service. It replicates data from SaaS applications, databases, files and events into destinations such as cloud warehouses and lakes using more than 700 prebuilt connectors (per its pricing page), and also runs transformations (Quickstart data models and hosted dbt projects) and reverse ETL, sold as Activations, following its 2025 agreement to acquire Census. Fivetran completed its merger with dbt Labs on 1 June 2026; the combined company operates as "Fivetran + dbt Labs". It is priced by usage, measured in Monthly Active Rows (MAR).

Build versus buy. Airflow gives you the framework to build pipelines; Fivetran sells finished connectors. Airflow can schedule your own extraction code, and Fivetran has its own scheduler, so each can work without the other, but they solve different parts of the problem. For an open source ingestion alternative to Fivetran, see Airbyte vs Fivetran and Apache Airflow vs Airbyte.

Side by side

Aspect	Apache Airflow	Fivetran
What it is	Workflow orchestrator for Python-defined DAGs	Managed ELT service with prebuilt connectors, transformations and reverse ETL
Who writes the extraction	You (Python tasks, provider operators or another tool)	Fivetran (connectors maintained by the vendor)
Licence	Apache-2.0 open source	Proprietary SaaS
Hosting	Self-hosted or managed (Astronomer, Amazon MWAA, Google Cloud)	Fivetran-hosted service
Scheduling	Cron, timetables, asset-based triggers, manual and API runs	Per-connection sync frequency (15 minutes on Standard, 1 minute on Enterprise and above)
Cross-system dependencies	Core purpose	Transformations can run after their source syncs (integrated scheduling); nothing beyond Fivetran
Pricing model	Free software; you pay for infrastructure or a managed service	Usage-based on MAR per connection; free plan; paid plans
Integration with the other	Fivetran operator and sensor in a separate provider package	Sync can be triggered through the Fivetran REST API
Main trade-off	No licence fee and full control, but you build and maintain extraction code and the platform	Little to build or run, but cost grows with changed rows and you depend on the vendor's connectors

Key differences

Building pipelines versus buying connectors

With Airflow alone, each source becomes code: authenticate, page through the API, track what changed since the last run, map types, handle schema changes and rate limits, then load. Airflow provider packages help for many systems, but you own the logic and its maintenance when an API changes. In return there is no per-row fee, and you can handle sources no vendor supports.

Fivetran sells that work as a service. You authorise a source, choose a destination and Fivetran keeps the tables up to date, including adding new columns as the source changes. Fivetran's documentation describes its model as ELT: raw data is loaded and stays available alongside transformed data. The trade-off, in our view, is predictable engineering effort against a usage bill that grows with data change, plus dependence on the vendor's connector coverage and roadmap.

The Fivetran Airflow provider: publisher and status

There is no Fivetran provider in the Apache Airflow project's own provider list. The maintained integration is airflow-provider-fivetran-async, an Apache-2.0 package published in Astronomer's GitHub organisation (feedback goes to an Astronomer address), which Fivetran announced as built with Astronomer. Version 2.4.0 was released on 17 April 2026; its changelog records fixes for Airflow 3 compatibility and adds Airflow 3.0, 3.1 and 3.2 to its test matrix, and its metadata requires Airflow 2.9 or later. The older airflow-provider-fivetran package (last release 1.1.4, January 2023) lives in a repository under Fivetran's organisation that is now archived, so do not use it for new work.

The package provides FivetranOperator, which starts a sync for a connector and by default waits for it to complete in deferrable mode (polling on Airflow's triggerer rather than holding a worker slot), and FivetranSensor, which waits for a sync that Fivetran's own scheduler started. The Airflow connection uses your Fivetran API key and secret. A minimal example, based on the README (the connector ID is a placeholder):

# Python, Apache Airflow 3 with airflow-provider-fivetran-async 2.x
from datetime import datetime

from airflow.sdk import DAG
from fivetran_provider_async.operators import FivetranOperator

with DAG(
    dag_id="fivetran_salesforce_sync",
    schedule="0 3 * * *",
    start_date=datetime(2026, 10, 1),
    catchup=False,
) as dag:
    sync_salesforce = FivetranOperator(
        task_id="sync_salesforce",
        fivetran_conn_id="fivetran",     # Airflow connection with API key and secret
        connector_id="my_connector_id",  # from the connector's settings page
        wait_for_completion=True,        # the default
    )
    # downstream transformation and test tasks follow sync_salesforce

If you orchestrate syncs from Airflow, consider whether the connector should also keep its own Fivetran schedule. The README describes FivetranSensor for the case where Fivetran schedules the sync and Airflow only waits for it.

Scheduling: when Fivetran alone is enough

Fivetran runs each connection on a sync frequency set by plan (the pricing page lists 15-minute syncs on Standard and 1-minute syncs on Enterprise and Business Critical). Its transformations support integrated scheduling, which runs models automatically after the connector data they depend on is updated, as well as frequency-based schedules and the schedule in a dbt project's deployment file. If your pipeline is "load these sources, then build dbt models on them", Fivetran can cover it without a separate orchestrator.

Airflow becomes useful when work extends beyond Fivetran: steps on other platforms (Spark jobs, machine learning training, file drops, API calls), dependencies between pipelines owned by different teams, custom alerting and retries, or a single place to see every run. Then Fivetran becomes one task in a larger DAG.

Operations and governance

Running Airflow yourself means operating the scheduler, API server, DAG processor, triggerer, workers and a metadata database, and planning upgrades such as the move from Airflow 2 to 3. Managed services reduce that work; Amazon MWAA lists Airflow 3.3.1, Google Cloud's managed service lists 3.3.1 builds, and Astronomer's Astro Runtime 3.3 ships Airflow 3.3.

Fivetran runs the infrastructure. For sensitive data it documents data blocking, which excludes tables and columns from syncs so that personal data never reaches the destination, and column hashing, which hashes values with a per-destination salt before writing them, keeping hashed columns joinable. Per the pricing page, the Business Critical plan adds customer-managed encryption keys and PCI DSS Level 1 certification. With Airflow, equivalent controls are whatever you build into your tasks and infrastructure.

Pricing and licensing

Apache Airflow is free (Apache-2.0). You pay for the servers or managed service it runs on and for the engineering time to build and maintain pipelines. Managed Airflow providers each have their own pricing model; check their pricing pages or calculators.

Fivetran bills connections by Monthly Active Rows: rows inserted or updated in a calendar month, excluding unchanged rows and initial syncs, with each connection on its own cost curve. Listed on the vendor's pricing page in October 2026: a Free plan with 500,000 MAR for connections and 5,000 monthly model runs for transformations; Standard, Enterprise and Business Critical paid plans; a USD 5 base charge on standard connections with between 1 and 1 million MAR in a month; and a 14-day free trial for each new connection. Transformation runs above the free 5,000 per month are priced per run. Because cost depends on how many rows change in each source, model it on your own data; we do not estimate bills.

Pricing checked on the vendors' official pages on 7 October 2026. Prices change; confirm before buying.

Where each one leads

Apache Airflow strengths

Free and open source under Apache-2.0, with several managed offerings
Orchestrates any workload, not only ingestion, with dependencies, retries and run history
Handles custom or unsupported sources with your own code
Pipelines live in Git as Python and can be reviewed and tested

Fivetran strengths

More than 700 prebuilt, vendor-maintained connectors
Fully managed: no extraction code or infrastructure to run
Handles schema changes and incremental updates automatically
Built-in transformations with integrated scheduling, plus data blocking and column hashing
Free plan for low volumes

Limitations

Apache Airflow limitations

Does not extract or load data by itself; you write or install the code
Self-hosting requires operating several components and planning major upgrades
Python skills are needed to author DAGs
No official Fivetran provider in the Apache project; the integration is a third-party package

Fivetran limitations

Usage-based cost grows with the number of changed rows per connection
Proprietary service; connector coverage and behaviour are the vendor's choices
Orchestration stops at Fivetran's own syncs and transformations
Faster sync frequencies require higher plans

When to choose each

Choose Apache Airflow if

You need to coordinate many systems and steps, not just loading data
Sources are custom, internal or not covered by a connector vendor
Licence cost matters more than engineering time
You already run Fivetran and need something to drive it alongside other work

Choose Fivetran if

Most sources are common SaaS applications or databases that Fivetran supports
The team is small and would rather not maintain extraction code
Load-then-transform with integrated scheduling covers the pipeline
You want data blocking and column hashing for personal data without building them

When neither is right

You want prebuilt connectors but prefer self-hosting or a different pricing model: see Airbyte vs Fivetran and the Fivetran alternatives guide.
You want orchestration with a different programming model: compare Airflow vs Dagster and Airflow vs Prefect.
All your data already lands in one warehouse and only transformation is missing: see ETL vs ELT.

Final recommendation

Bottom line

If the bottleneck is getting data out of common sources, Fivetran removes the most work, at a usage-based price you should model on real volumes. If the bottleneck is coordinating many steps and systems, or your sources are bespoke, Airflow is the more flexible and cheaper-to-license foundation, at the cost of engineering effort. When you need both, use Fivetran for ingestion and drive it from Airflow with the Astronomer-published async provider, not the archived original package.

Frequently asked questions

Is Fivetran a replacement for Airflow?

Not in general. Fivetran schedules its own syncs and can run transformations after them, which covers simple load-then-transform pipelines. It does not orchestrate work in other systems, so teams with broader pipelines usually keep an orchestrator.

Who publishes the Fivetran provider for Airflow?

The maintained package, airflow-provider-fivetran-async, is published in Astronomer's GitHub organisation under Apache-2.0 and was announced by Fivetran as built with Astronomer. It is not part of the Apache Airflow project's own providers. The older airflow-provider-fivetran repository under Fivetran's organisation is archived.

Does the Fivetran provider support Airflow 3?

Its changelog for 2.4.0 (April 2026) includes fixes for Airflow 3 compatibility and adds Airflow 3.0 to 3.2 to its test matrix; the package requires Airflow 2.9 or later. Check the changelog for the Airflow version you run.

What does Fivetran charge for?

Monthly Active Rows per connection (rows inserted or updated in a month), plus transformation model runs above a free monthly allowance. See the dated pricing section above for the plan details listed in October 2026.

Can Airflow load data without a tool like Fivetran?

Yes. Tasks can run Python extraction code or provider transfer operators. You then maintain each integration yourself, including incremental logic and schema changes.

Sources

Checked October 2026.

How we research comparisons: our editorial method.