Backups have moved from a background chore to a contested front line. Attackers now go after recovery systems before they touch anything else, regulators have started to ask for evidence of testing, and platform teams run workloads whose data does not sit in a tidy server any more. This overview of where BackupOps stands today (a label the Ops Atlas calls informal) summarises the trends the Atlas documents and the questions they raise for engineering organisations, including those in Japan and elsewhere in Asia that are building out cloud and container estates.
From copies to proof: the shift in what is being measured
The most important trend is conceptual. Older practice counted copies and schedules; the newer view counts evidence. The Atlas compresses the idea into one line: a copy earns trust once a test has brought it back. It then proposes matching indicators, among them asset coverage, adherence to loss targets, the age of the latest restore test, isolated-copy coverage for top tiers, restore duration compared with budget, and one combined score for proven recoverability. It describes plain job success rate as a vanity metric. Whether the vocabulary spreads or not, the question behind it, “have we actually done this?”, is becoming harder for management to dodge.
The rule of thumb keeps growing
The three-copies, two-media, one-offsite rule has a long history. The Atlas attributes it, via a 2012 US-CERT paper, to a 2009 book on managing digital assets. In 2019 the UK’s national cyber security centre added an offline element, and vendors later stretched the pattern into 3-2-1-1-0, which add an unalterable or offline copy plus a goal of zero verification errors. The pattern of expansion tells its own story: each extra digit answers a way that earlier arrangements failed. The Atlas suggests counting independent failure domains instead, which cuts through the numerology.
Threat pressure on backup systems themselves
Recovery infrastructure is now a target. A survey that the Atlas cites, run by a security vendor in early 2024 among 2,974 IT staff, found 94 percent reporting that attackers had tried to affect their backups, with 57 percent of those attempts reported as successful. These are self-reported figures and should be read as such. More concretely, the entry observes that the US government’s catalogue of exploited vulnerabilities, as of 16 October 2026, carried thirteen or more entries for backup software from a range of suppliers, and seven of those had been used in ransomware attacks. The implication for operations teams is unglamorous: patch the backup servers promptly, and keep them off the general-purpose network.
Regulators ask for rehearsal, not intention
For financial firms in the EU, Article 12 of DORA wants backup procedures on paper, exercised at intervals, with recovery systems segregated from the live estate. In the United States, a 2025 proposal affecting health-sector security rules envisaged monthly sample restores and short restore times, though reports suggest it will not be final before July 2027. Guidance from the UK’s national cyber security centre, published in November 2024, sets out principles for ransomware-resistant backups. Taken together, the direction is clear: documentation of policy is no longer enough, and organisations are expected to show what happens on the day of a restore. Readers in regulated sectors should consult their own advisers about what applies to them.
Cloud vaults, locks and their sharp edges
Managed vault features have made isolation easier to buy. The Atlas describes one cloud provider’s backup service where compliance mode locks the vault for good once a grace window of 72 hours or more has passed, whereas governance mode can be overridden by someone holding the right permission. The lesson it draws is about cost as much as safety: locked retention mistakes persist, and compliance-mode locks cannot be undone. Closing an account can also delete its contents after 90 days. Teams are therefore treating lock settings as change-controlled decisions rather than defaults.
Containers and SaaS: where the data actually lives
Two environments challenge the old server-centric model. In Kubernetes, the Atlas examines the open-source Velero project, now in the CNCF Sandbox: its hooks time out after 30 seconds, it lacks native support for immutable storage, and the credential it uses needs delete rights, which conflicts with the advice to keep one copy beyond reach. In hosted SaaS, the customer remains responsible for the data even when the supplier offers a built-in backup service that lives within the supplier’s own boundary. Both areas are growing fast, and both are where classic assumptions break first.
What careful teams do differently
The patterns the Atlas describes are modest and repeatable. Maintain a register with an owner per asset. Alert on the absence of success, and route those alerts outside production channels. Restore a sample on a schedule, then a full sandbox rebuild, then rehearse under time pressure. Keep encryption keys in separate escrow. These habits sit alongside the wider shifts in an essay on how DevOps practice may evolve in the Japanese market and beyond, and alongside the people side covered in a hands-on guide to culture change management, since recovery testing fails far more often for lack of ownership than lack of tooling.
Open questions worth watching
Several things remain unsettled. Nobody yet agrees on a single measure of proven recoverability, so comparisons between organisations stay approximate. It is unclear how far hosted-service providers will expose independent copy options to customers. And the pull between erasure obligations and retention locks is far from resolved; the Atlas advises keeping an erasure log that is re-applied after any restore, which is a stopgap more than a solution. Expect these debates to shape the next year of practice.
Reading the original
Most figures, dates and cases summarised above are taken from the long Ops Atlas page about BackupOps, which also lays out a maturity ladder running from “hope” up to “proven”. As the Atlas concedes that this is no settled field, treat the trends above as a snapshot of practice rather than a formal standard, and check any date or figure at the source before relying on it.









