Companies that hand their servers to a managed-support partner often assume the partner has also taken over the question of what happens in a disaster. Sometimes that is true; often it is not written down anywhere. Drawing on the Ops Atlas entry for DisasterRecoveryOps, this guide separates what a support contract usually covers from what recovery actually requires, and shows how to test the gap with evidence instead of reassurance.
Start with the plain definition
Disaster recovery is the set of arrangements for getting an IT service working again after a serious disruption. Cloud providers describe it as a subset of the broader business-continuity effort, which also covers people, premises and suppliers. The Atlas adds a useful nuance from the US government’s own guidance: its formal disaster recovery plan applies to disruptions that require relocation, meaning moving the work somewhere else. Smaller outages belong to ordinary incident handling.
DisasterRecoveryOps, the Atlas says, is the label for running all of this as a continuing operational practice. It is a young term, but the questions it covers are easy to state: how fast must we be back, how much recent work can we lose, who decides to switch, and how do we prove it works?
Four questions a support contract should answer
- Who owns the recovery objectives? A support partner can propose numbers, but only the business can sign them off, because only the business knows what an hour of downtime costs.
- Who holds the authority to declare a failover? If the answer is “whoever is on call”, check that this person can be reached and is permitted to act.
- Who runs and records drills? The Atlas treats rehearsal as the difference between a plan and a hope.
- Who is paid to fix gaps found in a drill? Findings without budget become a backlog nobody reads.
These questions tie in with the broader principle that delivery practices should be explicit and shared, a theme developed in an overview of core principles for effective software delivery.
Why the decision time matters so much to a support team
Here is the idea the Atlas returns to repeatedly: the recovery clock starts at the problem, and it includes the human decision. Its worked example reports a decision that took 25 minutes against an allowance of 16, a nine-minute overrun, and a total recovery of 64 minutes against a 60-minute objective. The technical team did their part; the delay sat in deciding.
For a managed-support arrangement, this has a concrete implication. If your partner’s engineer must phone your manager, who must consult a director, who is travelling, the clock keeps running. A good contract therefore names a pre-authorised decision-maker, with a deputy, and sets a deadline measured backwards from the objective.
What the provider does and does not do for you
Keep two kinds of provider apart. Your cloud or hosting platform runs the underlying infrastructure. Your managed-support partner runs operations on top of it. Neither automatically owns your recovery design. Even Microsoft’s own documentation, as the Atlas cites it, says DR is not something Azure supplies on its own. The customer has to choose a design and pay for it.
Consider the 2021 Strasbourg data-centre fire described in the entry. One site was destroyed, and some customers discovered that their copies were stored in the same place. That is an arrangement problem, not a hardware problem, and a support partner is well placed to spot it by asking where copies live relative to the thing they protect.
Reading a recovery scorecard like an auditor
Evidence beats adjectives. The Atlas example scorecard contains five numbers worth understanding: a recovery time that missed its target by four minutes, the decision overrun above, a data-loss window of 2.5 minutes against a 5-minute allowance, 67 percent of in-scope services covered by a drill, and 25 percent conformance of the written plan to what actually happened.
The last two figures are the revealing ones. Coverage says how much of your estate has been tested at all; conformance says whether people followed the plan or improvised. A partner who reports both is giving you a view of reality, not a sales page.
Common ways managed recovery goes wrong
- Recovery steps that live in a console and in someone’s memory. The entry files this under “click ops”, an anti-pattern borrowed from AWS guidance.
- A fence that does not fence. Switching a DNS route off redirects only traffic that follows DNS; batch jobs and clients with cached addresses may carry on writing to the old side.
- Recovery tools that need the failed region to work. In AWS’s us-east-1 trouble of October 2025, STS and Identity Center sign-in broke for some users, hence the entry’s advice to keep holding recovery-only credentials outside the failure domain.
- Standby capacity that has gone stale. Quotas and machine images drift unless somebody checks them.
Administrators moving into this kind of work can follow the route laid out in an argument for ops engineers adopting code-driven infrastructure and SRE habits, a good companion for anyone scripting recovery steps.
A simple review you can ask your partner for
- The signed objectives for each service, with the date they were last reviewed.
- The most recent drill report with time to decide, time to recover and data-loss window.
- The list of stores each critical service depends on and the replication lag of each.
- The failback plan, since the way home is a migration with its own hazards.
- Evidence that break-glass access was tested this quarter.
If the partner can produce those five items in a day, you are in good shape. If it takes a month, you have learned something valuable at a cheap moment. Clarity about who does what matters here as well, and the guide to DevOps role titles and the work behind them can suggest which job descriptions to name when you request a drill.
Where to read the primary source
For the full treatment, including maturity levels and metric definitions, see the Atlas entry that defines DisasterRecoveryOps and its scope. Its length is considerable, so begin with the decision-clock and anti-pattern sections.
Summing up
A support contract is not a recovery plan, and a recovery plan is not a drill result. Each step turns a claim into something you can check. Ask for the numbers, measure the decision, and test the way back. The cost is a few hours of engineering time each quarter; the benefit is knowing what will actually happen on the worst day.