Build Series Vol. 05 — Backup & Recovery

Copies an attacker holding domain admin cannot reach, and restores you have actually performed and timed — including the full-estate rebuild nobody plans for.

The decision this volume is really about: honest recovery objectives, and immutability that credential isolation makes real.

Control mapping first. The exact SOC 2, ISO 27001:2022, NIST CSF 2.0, CIS v8.1, CMMC, PCI DSS v4.0.1 and HIPAA clauses this build satisfies are printed before any installation step, so the work becomes audit evidence rather than a side effect of it.

A validation harness, not just a build. How to prove the control works and — the part nothing free covers — how to catch it failing quietly: the expired scan credential, the canary that stopped, the encryption key nobody can recover, the segmentation your diagram claims and your firewall does not.

Three sized architectures with honest hour counts. Tier 1 is 20–35 hrs to build · 3–5 hrs a month.

Stack: Restic · Borg · MinIO · Proxmox Backup Server · object-locked storage.

Current as of August 2026. Restic reads its repository index on every run, so it cannot be write-only — the correct pattern is a deny-delete credential policy, not a read-only one. And retention shorter than attacker dwell time means every backup you hold contains their access.

Every volume runs the same nine sections: prerequisites, control mapping, the decisions that matter, three sized architectures, a build runbook, a validation harness, an evidence pack, an operating cadence with honest hour counts, failure modes and the buy line, and an AI prompt pack.

Reading order: read 4th of ten. Each volume stands alone, and each is cheaper to build if the one before it exists.

Ships without support. This is a document, not a service contract — and Section 8 tells you plainly where the buy line is.