Uploaded November 2025 | Updated September 2026, 3 weeks ago
Could Ceph Become More Like a Database? - Radoslaw Zarzynski, IBM
For years, the RADOS layer of Ceph has provided key-value facilities known as OMAP. This feature is now extensively used by higher-level services like RadosGW and CephFS to store metadata for their objects (e.g., S3 objects, files & directories) and to implement custom indexing solutions. However, a crucial question remains: Can RADOS itself do more to enhance the OMAP experience?
This presentation provides an update on the research into so-called "RADOS Native Indexing." Currently, clients are responsible for their own indexing, which incurs a significant penalty in both performance and complexity. This cost originates from the lack of atomicity for operations spanning a data object and its corresponding index object, forcing clients to implement complex, multi-step commit protocols. Native indexing aims to eliminate this overhead by performing the index update atomically with the data object update, directly at the OSD level.
The author hopes that the subsequent Q&A session will help answer the title question and explore whether concepts like native indexing can spark new ideas and perhaps benefit the builders of next-generation distributed databases.
Could Ceph Become More Like a Database? - Radoslaw Zarzynski, IBM
For years, the RADOS layer of Ceph has provided key-value facilities known as OMAP. This feature is now extensively used by higher-level services like RadosGW and CephFS to store metadata for their objects (e.g., S3 objects, files & directories) and to implement custom indexing solutions. However, a crucial question remains: Can RADOS itself do more to enhance the OMAP experience?
This presentation provides an update on the research into so-called "RADOS Native Indexing." Currently, clients are responsible for their own indexing, which incurs a significant penalty in both performance and complexity. This cost originates from the lack of atomicity for operations spanning a data object and its corresponding index object, forcing clients to implement complex, multi-step commit protocols. Native indexing aims to eliminate this overhead by performing the index update atomically with the data object update, directly at the OSD level.
The author hopes that the subsequent Q&A session will help answer the title question and explore whether concepts like native indexing can spark new ideas and perhaps benefit the builders of next-generation distributed databases.
