Retention, immutability and separate copies
Retention, immutability and separate copies: why a deletable copy is not a backup.

A copy is written once; it becomes a backup only if it survives the event that destroyed the original. Survival is not decided by the transfer — replication, export, snapshot, job — but by four properties of where the copy lands: still there when recovery is attempted (retention), not rewritable in place (immutability), beyond the reach of a fault or an attacker on the primary (separation), and deletable only through credentials the primary’s compromise does not hold (control). A copy that fails any one is still a second copy of the data, and the same event that took the original can take it too.
The model: four properties of a copy that survives
┌───────────────┐ a copy survives only if all four hold
│ primary │
│ data + cre- │ 1 retention still present when recovery is attempted
│ dentials A │──────▶ 2 immutability cannot be overwritten or deleted in place
└───────┬───────┘ 3 separation a fault on the primary does not reach it
│ 4 control credentials A cannot delete it
an attacker who owns A
deletes, encrypts or
overwrites every copy
it can reach
The four are independent, and none substitutes for another: immutability without retention protects a closed window, retention without separation keeps a copy the same fault can erase, separation without control keeps a copy the primary’s administrator can delete.
Retention is a policy: how long, and why
SNIA calls the retention period “The length of time that a backup copy should be kept”, and NIST notes that “Data retention is usually satisfied by keeping a copy of data on some backup medium” — for operational, legal or regulatory reasons. The force is in should: retention is a decision about a duration, answered with a reason, not a storage figure.
For this sheet’s failure one reason dominates: the window must outlive detection. NIST describes an attacker who lets the copy age out — “When existing copies cannot be compromised, another viable attack strategy could be to interfere with the backup process itself” — then returns “knowing that the only available copies for recovery are too old”. A daily copy kept seven days answers a fault noticed that day, nothing an intrusion found on the twelfth. The test: does the window exceed the time between a compromise and its discovery?
Immutability: WORM, object lock and legal holds
Immutability is the technical form of a copy that cannot be rewritten. SNIA calls WORM storage one “designed for fixed content, that preserves what is written to it in an immutable fashion”; NIST states it plainly: “Immutability involves the ability to lock data after it has been created, thereby preventing it from alteration or deletion”. The canonical form is object lock on a versioned object store — Amazon’s S3 Object Lock, a “write-once-read-many (WORM) model to store objects”.
A retention period “specifies a fixed period of time during which an object version remains locked” — fixed, when the retain-until date is set at write time, or variable, when an event sets it on release: Amazon’s “configurable recovery window against ransomware or accidental deletion”. A legal hold has no expiration and “remains in place until you explicitly remove it”; a default bucket retention protects every object that lands in it.
The mode sets the strength. In compliance mode a locked version “can’t be overwritten or deleted by any user, including the root user”, and the only way out before the date “is to delete the associated AWS account”; in governance mode a user “can’t overwrite or delete an object version or alter its lock settings unless they have special permissions”, a bypass right that makes governance the weaker mode.
Where immutability stops
- The window ends. After the retain-until date, “the object version can be overwritten or deleted” like any other. Immutability is a lease, not a guarantee.
- It protects a version, not the object. A lock does not stop a new version being written over it: “Retention periods and legal holds don’t prevent new versions of the object from being created”.
- It is a control on the API, not on ownership. Locking blocks the delete and overwrite calls; it does not remove the credential that can change the lock or the account that can be deleted.
Separate copies: medium, fault domain and control
Separation answers reach. NIST’s backup control requires an organization to “Protect the confidentiality, integrity, and availability of backup information” and to store critical backups “in a separate facility or in a fire rated container that is not collocated with the operational system”. On the medium, this answers a correlated fault: a disk, host, rack or site that takes the primary must not take the copy.
Separation of control answers the attacker who owns the primary. NIST requires that “servers and clients should not be allowed to change their own data protection configuration”, and puts the privilege to “configure, stop, and delete backup” in a role separate from data management. Offline storage is the strictest form: NIST advises organizations to set up “an air gap around the cyber-attack recovery copies of sensitive data”, and in the strict case air-gapping “should provide full physical and network-level separation”. Offline is thus a control on credentials, not only on the medium — a copy whose delete path needs a credential absent from the primary cannot be deleted by the attacker who owns it. CISA states the practice and its cause: organizations must “maintain offline, encrypted backups of data”, because “many ransomware variants attempt to find and delete any accessible” copy.
The 3-2-1 heuristic and the common error
The shorthand for this is 3-2-1 — three copies, on two kinds of media, one off-site. It is a useful checklist because it forces a copy off the primary’s fault domain, and that is all: a convention, not a definition any source here states. It describes where the copies are and says nothing about whether they can be deleted, rewritten or reached with the primary’s credentials. NIST sets the requirement properly when it asks for “immutability and locking, minimum number of copies per backup set, and the geographic distribution of such copies”.
The common error follows. An administrator restores from a second copy on another host; if one administrative identity can delete both, one account still owns every copy, and the copy has added a fault domain but not a control. A copy is a backup when it is retained past detection, immutable for that window, on separate storage and under separate control.
Level and prerequisites
L3 — the properties of a surviving copy, not the procedure that produces it: retention as a policy, immutability as a lock, separation as a control on medium and credentials. Prerequisites: the block, file and object sheet; full, incremental and differential backup; RAID and backup; snapshots and their limits; replication and the mistakes it replicates.
Where to go next
- Server & Virtualization — the area this sheet belongs to.
References
- SNIA, SNIA Dictionary — retention period, data retention, WORM storage and non-erasable content.
- NIST, Security Guidelines for Storage Infrastructure (SP 800-209) — immutability, data retention, air gap, immutable storage for recovery copies, separation of duty.
- NIST, Security and Privacy Controls for Information Systems and Organizations (SP 800-53 Rev. 5) — CP-9 System Backup, its separate-storage and deletion-authorization enhancements.
- CISA / MS-ISAC, Ransomware Guide — offline, encrypted backups and why they must be offline.
- Amazon Web Services, Locking objects with Object Lock — the WORM model, retention periods, legal holds and retention modes.