Kent Overstreet tagged version 1.38.6 of bcachefs-tools this week. He calls it the performance release. And with it comes a clear message. The filesystem once flagged as experimental no longer carries that label.
Overstreet dropped the experimental tag some time ago. “Consider this the belated official announcement :),” he wrote in his Patreon post. The announcement arrives months after the project left the mainline Linux kernel. Distributions now ship it as a DKMS module. That shift followed friction with kernel maintainers. Yet development continued. Faster than before in some respects.
The new tools support up to 255 devices per filesystem. They fix half a dozen bugs. Six performance optimizations stand out. Journal flushes run fully lockless. Per-pin-list locks reduce contention. Binary search speeds journal reads. Btree sharding now uses PID-based logic with lock migration. Multithreaded write-buffer flushing joins the list. Reconcile operations run quicker and with more parallelism. Erasure coding graduates too. “In use and seems to be working quite well,” Overstreet noted.
These changes matter for real workloads. On an Epyc 9454 with 48 Zen 4 cores, bcachefs pushed 16.5 GB/sec through dbench with 48 clients. XFS reached 16 GB/sec on the same hardware. The Register highlighted the numbers. Four-kilobyte random writes with fio delivered 700,000 IOPS for bcachefs against one million for XFS. The filesystem performs extra verification and checksumming. Those figures feel more believable given the added work.
Phoronix reported the release two days ago. It pointed to heavy focus on journalling code. Lock contention dropped sharply. The updates introduce no on-disk format changes. Users can adopt them without fear of incompatibility. Ubuntu 26.04 LTS already offers packages through apt.bcachefs.org. Other distributions provide builds through their channels or COPR repositories. Phoronix noted the steady stream of point releases this year. Version 1.37 arrived in March with stable erasure coding and Linux 7.0 support.
Bcachefs itself traces roots to the original bcache block cache. Overstreet designed it as a modern copy-on-write filesystem. Features include checksumming for data and metadata, replication, compression, encryption, snapshots, and scrub. It scales to petabyte sizes. Online fsck remains a work in progress. The project website lists these capabilities plainly. It also explains the move to DKMS after kernel 6.18. The in-tree code had grown stale. Continued inclusion risked confusion. bcachefs.org directs users to distribution packages for easy installation.
Personality clashes played a role in the kernel exit. Overstreet and Linus Torvalds disagreed on development pace and patch discipline. The filesystem was removed from mainline. Some observers called the episode unfortunate. Overstreet expressed hope it might return one day. Perhaps as the first major filesystem written in Rust.
That transition has begun. Userspace code already runs in Rust. Safe Rust interfaces exist for the core btree iterator and utility functions. The next release will pull these into the DKMS module. Core kernel code conversion follows. Overstreet expects roughly 50 percent Rust this year. Converting the journal to safe Rust will prove interesting. He acknowledges the challenge directly.
Overstreet also addressed artificial intelligence in coding. He holds no opposition to the tools. Yet he sees lazy submissions. “I’m definitely not against AIs, but I am starting to see the lazy patch submissions/bug reports where someone clearly is just asking an LLM to do all the work, and that’s not ok,” he wrote. The comment reflects broader tensions in open-source maintenance.
Performance work accelerated after the kernel departure. Earlier versions targeted feature parity. Now the focus sharpens on speed. Journal and allocator changes in 1.38 reduced mount stalls. Discard handling improved under load. Systems with many snapshots mount faster after an O(n²) issue was fixed. These incremental gains compound.
Independent tests show mixed results against established filesystems. Bcachefs sometimes trails XFS or ext4 in raw throughput. It often leads in features. Snapshots, inline compression, and per-file encryption come without separate layers. The design avoids the complexity found in other advanced filesystems. That coherence appeals to certain users.
Adoption grows despite the experimental history. Arch Linux and NixOS include packages by default in some editions. Debian and Ubuntu users add the apt repository. Fedora taps COPR. OpenSUSE builds appear on community repositories. An experimental NAS operating system called NASty builds around bcachefs. Interest persists in storage communities.
Questions remain about long-term support. Without mainline inclusion, vendors bear more responsibility for maintenance. DKMS eases deployment but adds a compilation step. Bugs surface less often now. Overstreet maintains active development. The Patreon updates keep supporters informed. Recent preliminary benchmark posts generated discussion on forums.
The Rust effort could change perceptions. Kernel developers show increasing comfort with the language. Full conversion might reopen merger talks. For now the project stands on its own. Performance numbers improve. Features solidify. Users report fewer corruption incidents than in early days.
Storage administrators weigh trade-offs. Bcachefs delivers checksumming and snapshots natively. It handles tiering based on device performance without complex configuration. Hot data migrates to faster drives automatically. Cold data rests on slower media. The approach promises efficiency gains in large arrays.
Overstreet continues refining. No major format changes in 1.38.6 mean safe upgrades. The lockless journal work targets high-concurrency environments. Databases and virtual machines benefit. Fsync-heavy workloads see lower latency. These details matter to enterprise deployments.
So the filesystem matures. It left the kernel. It shed the experimental tag. And it posts competitive numbers against battle-tested alternatives. The coming Rust integration adds another dimension. Observers will track how distributions respond and whether performance leads widen. For now Overstreet’s latest release signals confidence. The project moves forward on its own terms.