A Backup Target That Holds: Repository Server Hardware
Backup infrastructure gets specced last and blamed first. The repository server looks like the least demanding box in the room — it mostly receives sequential writes at night — right up until a restore is needed at 10 a.m. on a Tuesday, or a synthetic full consolidation grinds, or ransomware goes looking for it. Speccing it deliberately is cheap insurance, and refurbished 12LFF platforms are almost purpose-built for the role.
Capacity math, then rebuild math
Size the pool from retention policy: full-equivalent size × change rate × retention days, divided by the dedup/compression factor your backup software honestly achieves (2–3× is a safer planning number than vendor best cases). Then check the second number people skip: rebuild exposure. Twelve 16TB drives in one RAID6/RAIDZ2 group rebuild for a long time; two six-drive groups halve the window and the risk. On a 12LFF R740xd or DL380 12LFF, two six-wide groups is the sane default layout.
The CPU is not decorative
Modern backup engines do real compute on the repository: compression, dedup hashing, encryption, synthetic-full construction, health checks that read everything back. A single mid-range CPU (12–16 cores) with 64–128GB RAM keeps those jobs off the clock; the starving-the-repo pattern — old dual-core leftovers, 16GB — is why some synthetic fulls take all weekend. Refurbished platforms make the right answer nearly free.
Network: match the restore you promised
A 1GbE link moves roughly 400GB per hour under ideal conditions. If the recovery objective is a 4TB server in under two hours, the arithmetic already demands 10GbE end to end. Dual SFP+ on the repository, and — often forgotten — on the hosts being restored to, is what makes RTO numbers honest.
Immutability belongs in the design
The modern threat model includes an attacker who finds the backup server with admin credentials in hand. Answers that hold: a hardened Linux repository with immutability flags (the pattern Veeam and others formalize), object storage with versioning/object-lock as a second tier, and at least one copy that is offline or offsite by construction. Hardware-wise this argues for the repository being its own physical box with its own credentials — not a VM inside the cluster it protects, not a share on the production NAS. The 3-2-1 rule survives because it keeps being right.
A reference build
Dell R740xd 12LFF: 1× Silver 4114, 96GB RAM, HBA or RAID as your backup software prefers, twelve 12TB SAS drives as two six-wide RAID6/RAIDZ2 groups (≈96TB usable before dedup), a mirrored SSD pair in the rear flex bays for OS and metadata, dual 10GbE SFP+. Equivalent HPE: DL380 Gen10 12LFF with a P408i or HBA mode. Either box sits quietly for years — and on the day it matters, restores at the speed you actually planned for.