IT Brief UK - Technology news for CIOs & IT decision-makers
United Kingdom
croit unveils Rust-based Ceph FlashStore OSD preview

croit unveils Rust-based Ceph FlashStore OSD preview

Thu, 19th Nov 2026 (Today)
Mara Sugue
MARA SUGUE News Editor

croit has released a public preview of FlashStore OSD, a Rust-based Ceph storage daemon for NVMe flash that works with existing Ceph clusters.

The launch is the first public step in croit's broader effort to rewrite parts of Ceph in Rust while keeping the new components compatible with the standard Ceph protocol. Alongside FlashStore OSD, it has also released public previews of Rust client libraries for RADOS and RBD, as well as an RDMA transport.

Ceph is widely used as an open-source storage platform, particularly in large-scale environments, but users running demanding workloads have long complained about flash performance. croit's new design aims to reduce the CPU overhead of the classic Ceph object storage daemon, which was built when hard disks, rather than flash drives, were the main bottleneck.

In company tests on a 30-NVMe OSD cluster, FlashStore OSD delivered average throughput gains of 50 per cent on three-way replicated pools and 54 per cent on 2+2 erasure-coded pools, compared with stock Ceph v21.2.0 BlueStore. croit also reported that 99th-percentile tail latency roughly halved and CPU time per operation fell by 36 to 37 per cent.

The benchmark setup used five AMD EPYC 9455P servers, each with six OSDs and 2 x 200 Gbit/s networking, using standard kernel TCP rather than RDMA. According to croit, the biggest gains came in workloads where the classic OSD spends the most CPU time per operation, including up to 123 per cent higher throughput for 4K sequential reads on replicated pools and up to 187 per cent higher throughput for 4K random writes on erasure-coded pools.

At lower queue depths, the gains were narrower. Replicated IOPS at queue depth 1 were 9 per cent higher on average. Across 28 tests, FlashStore OSD matched or exceeded BlueStore throughput in 27, with one 64K random read test on replicated pools described as effectively level.

Architecture change

To the rest of the cluster, FlashStore OSD appears as a standard Ceph OSD, meaning it can run alongside existing monitors, clients and traditional OSDs. Internally, however, the design differs in several ways.

Each CPU core owns its share of the data, so threads do not wait on one another's locks. The software also replaces RocksDB with a new storage engine, uses io_uring for drive and network access, and is designed to minimise data copying as writes are passed to drives and replicas. A write is acknowledged only after it has been stored on every replica.

croit also said memory use remains more predictable than with BlueStore, with median usage of 2.2 GiB to 4.8 GiB per OSD. It contrasted this with BlueStore's use of RocksDB metadata in the kernel page cache, which can grow as the data set expands.

For key-value metadata known as OMAP, used by the S3 bucket index and CephFS, FlashStore OSD was 22 per cent faster on average in tests run through RADOS, according to croit. The company did not publish benchmark figures for the S3 gateway or CephFS themselves.

“Performance is the objection we hear most often about Ceph, especially from customers moving off VMware or building AI workloads,” said Andy Muthmann, Chief Executive Officer, croit. “With croit FlashStore OSD we get considerably more out of the hardware that is already there. Anyone running a flash cluster today gets more performance without buying new servers.”

Broader rewrite

FlashStore OSD is part of a wider Rust stack that croit is building around Ceph. The related preview releases include librustyrados, described as a drop-in replacement for Ceph's librados with a native Rust interface, and librustyrbd for RBD block devices used by QEMU/KVM.

croit said the stack is not a fork of Ceph because the new components remain wire-compatible with existing clusters and can be deployed gradually. That means operators can migrate one drive or one server at a time while classic and new OSDs continue serving the same pool.

“We have run Ceph in hundreds of customer clusters for almost ten years, and we know where it hurts on NVMe,” said Martin Verges, Chief Executive Officer, croit. “croit FlashStore OSD is built for today's hardware: one thread per core, no locks on the hot path, its own object store. Less overhead per operation and lower latency, on top of a better architecture, is what gets more performance out of the drives our customers already own.”

FlashStore OSD can run on Ceph v21.2.0, v20.2.5 and v20.2.4, even without croit's own management software. It requires NVMe SSDs with power-loss protection, a 4 KiB sector format, and Linux 5.14 or newer with io_uring enabled.

The software is being released as open source, with FlashStore OSD under LGPL-3.0 and the Rust client libraries under MPL-2.0. croit intends to offer its work to the upstream Ceph project, although any decision on merging it into the main project will rest with the broader community.

The engineering team behind the work includes Igor Fedotov, a BlueStore developer and member of the Ceph Steering Committee, and Md Mahamudur Rahaman Sajib. croit said its developers use AI as a tool during development, while design, review and testing remain with the engineers, and each change is benchmarked against BlueStore.

“We want Ceph to stay the right choice for the next twenty years, not just the last twenty,” said Verges. “That is why we are rewriting it in Rust, wire-compatible, so the new components drop into existing clusters one drive or one server at a time. We publish them as open source, together with the client side, so the whole Ceph community can try them, review them and build on them.”

FlashStore OSD is available for testing and evaluation and is not yet recommended for production data.