Apache Airflow is the open-source scheduler most data teams use: Airflow 3.0 shipped on April 22, 2025, 3.1 added human-in-the-loop tasks and deadline alerts on September 25, 2025, and Airflow 2 reached end of life on April 22, 2026 [1][2][3]. Astronomer runs Airflow as a managed service, Astro [5]. Lima Backbone runs on Airflow: Lima tasks appear inside your DAGs. So this is a comparison of what a scheduler does and what an engine does, and of what Lima adds when the two run together.
Each row states what each product does in plain words. Where a fact depends on a vendor document, the document is numbered and linked beneath the table.
| Item | Apache Airflow and Astronomer | Lima Backbone |
|---|---|---|
| What it is | Airflow: an open-source platform to author, schedule and monitor workflows as Python DAGs [1]. Astro: Astronomer's managed Airflow with hosted deployments and dedicated clusters [5]. | Lima Backbone: the data-movement part of the Lima platform. Pipelines, orchestration, change capture and write-back, from a nightly copy into a data lake to a full ERP-to-planning integration with results returned to the ERP. |
| Where it runs | Self-hosted anywhere; Astro hosted deployments, dedicated clusters, or Astro Private Cloud [5]. | Lima's cloud, your private cloud, or on your own machines; the same software in each (product). |
| How you pay | Airflow is free software. Astro: Developer deployments from $0.35 per hour, Team from $0.42 per hour, workers from $0.13 per hour scaling to zero, dedicated clusters from $2.40 per hour; Business and Enterprise plans on request [5]. | A subscription sized by the tier your estate needs (32C to 128C); the whole platform is included; the price does not change with rows moved, connections, jobs or questions asked. Fixed-price entry offers (product; see the pricing page). |
| SAP extraction | Airflow schedules tasks; extraction is done by operators and provider packages or by other tools the tasks call. No packaged SAP application-layer extraction in Airflow itself [1]. | Table reads through a named released function module, BAPIs, released CDS views, change pointers and extractor delta queues. No ODP RFC (SAP Note 3255746), no RFC_READ_TABLE in production (Note 382318), no HANA log reading (Note 2971304). Nothing installed in SAP (product; compliance statement pending sign-off). |
| Oracle extraction | Oracle provider operators run SQL and transfers; change capture is done by the tools the tasks call [1]. | Partitioned JDBC reads within a cap agreed with the DBA; changed rows by last-update columns; LogMiner where the DBA allows; Oracle Fusion Cloud through BICC extracts and REST (product). |
| Change capture | Not a scheduler function; done by the called tools. | Application-layer change capture for SAP; database-log capture for PostgreSQL, MySQL, SQL Server and Oracle; streams in micro-batches (product). |
| Transformation | Python in tasks, or tools such as dbt and Spark called from operators [1]. | 150+ visual operators and Python on a parallel engine; ETL, ELT or a mix (product). |
| Orchestration and deadlines | Schedules, sensors, datasets and event-driven triggers; deadline alerts when a run exceeds a time; human-in-the-loop tasks since 3.1 [2]. | Runs on Apache Airflow with five trigger modes. A run window with a business deadline; a failed check pauses the run and its clock; resume from checkpoint; ordered replay (product). |
| A bad record | Task-level success, failure and retry; row-level handling is inside the called tools [1]. | Incoming = valid + exception records. A failed row is kept, routed to the buyer, planner or steward who owns it with a deadline, and replayed once fixed (product). |
| Write-back | Whatever an operator writes; no ERP write-back semantics in the scheduler itself. | To SAP through BAPIs and IDocs, to Oracle through open interface tables and public APIs; each write carries a key and counts as done only when the ERP confirms it; a ledger records every write (product). |
| Planning applications as targets | Through custom operators or other tools. | Kinaxis Maestro, o9, Blue Yonder, SAP IBP, OMP and Anaplan are packaged targets with their data models described (product). |
| Reconciliation | Task logs; row counts if a task records them. | A report after every run: rows in, rows out, control totals and variance per table; exceptions and their owners; deadline met or missed (product). |
| Standards | OpenLineage integration through the Airflow provider; the OpenLineage project originated around Airflow lineage [1]. | Airflow-native; OpenLineage, Open Data Contract Standard, MCP server, Iceberg and Delta output on a dated schedule (see the standards page). |
A tripped check pauses the run and the deadline clock, and the run resumes from its checkpoint. Airflow 3.1 alerts when a run exceeds a time [2]. Lima, running on Airflow, treats the business deadline as the promise: a failed quality check holds the dependent pipelines, the clock pauses with them, and the run resumes and replays in order once fixed. Test it: trip a check at 02:14 and watch the 06:00 load.
A comparison is only useful if it says when the other product fits. These are the situations in which each one does.
Lima runs on Airflow. An estate that already runs Airflow or Astro keeps its DAGs and adds Lima tasks to them; Lima's own Airflow can also host the DAGs. Airflow 3 is supported; Lima's statement on Airflow 3 is on the standards page (marked pending sign-off).
Estates on Airflow 2 have a second reason to plan: Airflow 2 reached end of life on April 22, 2026 [3]; Astro supports it through April 2027 [6].
No. Lima runs on Airflow. You keep your DAGs and add Lima tasks, or let Lima's Airflow run them.
Lima includes orchestration on Airflow with five trigger modes, business deadlines and pause, resume and replay. Its main job is the engine underneath: extraction, checks, transformation, loads and write-back.
Astro can host the Airflow that runs Lima tasks; the Lima engine runs where your data is.
A two-week bake-off on your estate measures rows per minute, hours to a first load, reconciliation variance, run-window adherence and exceptions routed, plus one confirmed write-back.