Bcachefs, an Introduction/Exploration

Original Article ↗ Hacker News Discussion ↗

Overall sentiment

  • Thread centers on comparing bcachefs, btrfs, and ZFS for reliability, features, and maturity.
  • Many are excited about bcachefs as a new in-kernel COW filesystem, but most are cautious about trusting it with important data yet.

Btrfs reliability and behavior

  • Multiple anecdotes of recent btrfs data loss or unrecoverable corruption (power cuts, full filesystems, “parent transid verify failed”, USB usage causing hard freezes).
  • Others report years to a decade of trouble‑free use, including large production deployments, and argue its bad reputation lags behind current reality.
  • Common pain points:
    • Fragility when the filesystem is 100% full.
    • RAID5/6 still officially unsafe; write‑hole problems.
    • Awkward management (balance/defrag behavior, snapshot/UUID quirks, drive removal with I/O errors).
    • Performance issues for databases/VMs due to COW; NoCOW seen as a hack that disables core features.
  • RAID1 behavior and defaults are controversial: refusing to boot without all devices is seen by some as safe, by others as an availability trap.

ZFS vs btrfs

  • Several users say ZFS has never lost data for them over many years and prefer it despite kernel/module friction.
  • Others note ZFS has had serious bugs too and is not invulnerable, especially without ECC RAM.
  • ZFS is praised for: robust snapshots, send/receive, checksumming, compression, flexible datasets, and as the backbone for backup schemes.
  • Criticisms: licensing/legal uncertainty on Linux; out‑of‑tree modules; dedup is available but considered expensive; encryption has open bug tickets but also long‑term positive anecdotes.

Bcachefs: promise and current limitations

  • Recognized as very new in‑kernel; many bugfixes landed right after merge.
  • Positives: unified feature set (COW, checksums, compression, encryption, dedup, SSD caching, flexible multi‑disk, per‑file replication). Some report smooth desktop use so far.
  • Negatives/concerns:
    • At least one report of it “eating data”; others only trust it for non‑critical systems.
    • RAID5/6 / erasure coding explicitly marked “DO NOT USE YET”.
    • RAID0‑like behavior by default with multiple disks worries some.
    • Even its advocates acknowledge they still keep important data on ZFS.
  • Recent change: upcoming Linux 6.11 adds self‑healing on checksum error; there is debate about risks of self‑healing with bad RAM, and the role of ECC.

RAID, multi‑disk, and uneven‑disk setups

  • Btrfs RAID1’s ability to mirror at the chunk level across different‑sized disks is considered a killer feature for small/home setups; bcachefs offers something similar.
  • ZFS can approximate this with manually partitioned vdevs but at the cost of complexity and performance.
  • General dissatisfaction that traditional RAID and most filesystems don’t gracefully handle disks that come and go; resilvering is seen as wasteful when little has changed, though ZFS’s incremental resilver is noted as relatively fast.