Open-source or commercial Spark accelerators: what actually changes

Apache Gluten, Apache DataFusion Comet and Apache Auron (formerly Blaze) are Apache Software Foundation projects under the Apache License 2.0. The RAPIDS Accelerator for Apache Spark, now published by NVIDIA as NVIDIA cuDF for Apache Spark, is also Apache 2.0. DualBird and Flarion are commercial products. Photon is commercial in a different way: it has no separate license because it is part of the Databricks Runtime, but Photon compute consumes Databricks Units at a different rate.

No license fee is not the same as no cost. The open-source plugins run on compute you pay for and on settings your team owns: off-heap memory size, the shuffle manager, spill limits and version upgrades. Gluten's Velox getting-started page shows an off-heap example of 20g, and Comet's tuning guide lists pool fractions, batch size and spill limits to set. The RAPIDS Accelerator has no plugin fee, but every node must be an NVIDIA GPU instance, so its cost is the change in instance price against the runtime saved. Spark cost per job, explained shows how to put these on one scale.

For the open-source projects, support is community-based: issue trackers, mailing lists and chat channels, with no vendor support contract described on the project pages we reviewed. Auron points users to the Apache mailing list. DualBird is vendor-supported, and Databricks manages Photon inside its runtime, so there is no plugin for your team to install or version. Flarion describes performance monitoring and anomaly detection but does not publish support terms. If a pipeline has a delivery deadline, write down the support path before the trial, not after the first failed run.

Governance tells you how a product is likely to evolve. Gluten became an Apache top-level project in March 2026 after starting at Intel and Kyligence in 2022. Comet is a subproject of Apache DataFusion and reached 1.0.0 on 7 August 2026. Auron is still incubating at the Apache Software Foundation. The RAPIDS Accelerator is a long-running NVIDIA project with more than 9,000 commits on main. The commercial vendors set their own roadmaps: DualBird announced $25 million in funding in February 2026, and Flarion does not publish company details such as founding date and funding on the pages we reviewed.

All seven products work under the Spark API rather than replacing it, so removing one leaves your Spark SQL and DataFrame code as it was. What you unwind is configuration and, for some, infrastructure: plugin jars and shuffle-manager settings for the open-source plugins, GPU instances for the RAPIDS Accelerator, F2 instances for DualBird, and a runtime setting for Photon, with the jobs staying on Databricks. Keep the baseline configuration from your trial so that switching back is a tested change.