The failure retention protects against
Fast failures (dropped table today) need recent backups. Slow failures — corruption or bad logic quietly mangling rows for weeks — need OLD backups, because every recent one contains the damage. Retention design is about the second kind.
GFS: 7 daily, 4 weekly, 12 monthly
Keep every daily for a week (fast failures), a weekly (say Sunday's) for a month, and a monthly (the 1st) for a year. ~23 archives cover you from 'oops just now' to 'this has been wrong since Q1' at a fraction of keep-everything storage.
Two hard rules
1. Never delete your most recent VERIFIED backup — age is irrelevant; it may be the only copy proven to restore.
2. Keep the audit trail after pruning — delete the object, keep the record (when taken, size, checks passed). Auditors ask about the policy's history, not just its present.
Firedrill enforces GFS automatically on Pro/Team and hard-codes rule 1.