Skip to content

Latest commit

 

History

History
30 lines (27 loc) · 3.28 KB

File metadata and controls

30 lines (27 loc) · 3.28 KB

Rejected Storage Defaults and Anti-Patterns

Project home | Documentation index | IEEE references

Rejected advice Reason Replacement
Ext4 for every workload Maturity alone does not provide checksums, snapshots, compression, or managed multi-disk recovery Select from the decision matrix
noatime,nodiratime everywhere Redundant and can break access-time consumers Keep default relatime; measure exceptions
lazytime everywhere Delays persistence of inode timestamps Use only for a measured workload that accepts the crash behavior
barrier=1 everywhere Already automatic where supported; obsolete on XFS Keep safe defaults and never use nobarrier generically
Ext4 data=ordered on hard disks only It is Ext4's default on every block-device type Omit it
Ext4 commit=30, 60, or 300 Enlarges the crash-loss window Keep the five-second Ext4 default
Btrfs compress-force=zstd:3 Wastes CPU on incompressible files; upstream no longer recommends forcing Use the Btrfs compression policy
Btrfs ssd,space_cache=v2,discard=async Current detection and defaults already select these behaviors when appropriate Omit them unless supporting an older measured target
Continuous discard on every SSD Can create latency and leaks allocation through encryption Use the filesystem-specific TRIM policy
fstrim.service as the scheduled unit It is the one-shot worker, not the scheduler Enable fstrim.timer
Disable flushes or barriers on Gen5 NVMe Link speed is unrelated to power-loss ordering Keep flush and FUA semantics
Force 4 KiB namespace or dm-crypt sectors The three block-size layers are independent; wrong topology can corrupt partial sectors Inspect and keep automatic defaults
Tune scheduler, queues, read-ahead, or dm-crypt workqueues by drive class Synthetic gains can worsen fairness, thermals, or real latency Benchmark the complete encrypted workload
Btrfs RAID5/6 for production Official status still identifies severe known problems RAID1, RAID1C3, RAID1C4, RAID10, or another storage design
Bcachefs as a universal 2026 production default Out-of-tree DKMS and recovery path remain fast-moving Follow the evaluation-only profile
exFAT as a backup or Linux data filesystem No journal, Unix permissions, or user-data checksums Ext4 plus LUKS2 for Linux backups; exFAT only for exchange
A fixed 8 GiB tmpfs for SSD longevity Consumes volatile memory or swap and changes failure semantics Use distribution defaults unless the workload needs tmpfs
Rebuild every initramfs after any fstab edit Most data-mount changes are not early-boot inputs Rebuild only for distribution-specific early-boot changes
Snapshots or RAID as backup Same-system deletion, compromise, controller failure, and operator error remain Independent encrypted, restore-tested copies

The corresponding safe defaults and exceptions are documented in mount options and TRIM, encryption architecture, and the filesystem profile index.