Btrfs/ZFS/bcachefs under workloads classic benchmarks skip

The discussion centers on a public filesystem benchmark authored by fenio, who explains that tests run on GitHub runners using loop devices with a calibration step to discard unreliable VMs and that he’s collected many runs and started limited real-hardware tests on Hetzner and SAS HDDs. Critics such as Farmadupe, magicalhippo, azatom and loeg challenge the reliance on shared ephemeral VMs, noisy neighbors, missing clear SSD/HDD baselines, and readability of results; walrus01 and koverstreet counter that a dedicated hypervisor or Hetzner machines could be reasonable for automated testing. Toast0, Joel_Mckay and others propose alternate setups (dedicated bare-metal partitions, additional filesystem types, networked storage scenarios) while fenio reiterates openness to improvements and notes the time cost of real-hardware runs.

Opinion splits sharply on technical and social grounds. Irusensei praises bcachefs for flexible caching, mixed-device support and per-file replication, while multiple commenters (sippingabonedry, blop) recommend ZFS for stability and express concern about maintainer drama and bus-factor risk - koverstreet and blop debate how much that should matter to users. Farmadupe and markhahn press for clarity on integrity failures and whether injected corruption was recoverable after scrubs; hlieberman points out integrity checks only apply to RAID-like tests. Overall the conversation divides between those who see CI-derived averages as useful signals and those who demand bare-metal, labeled baselines and clearer integrity validation.