Ask five engineers whether their system is covered for a disaster and you will hear five confident yes answers, each resting on a different assumption. Some of those assumptions are sound. Many are myths that survive because nobody has ever timed a real recovery. Here are six that come up again and again in DisasterRecoveryOps discussions, with the corrections the Ops Atlas offers for each.
Myth 1: “We run across several availability zones, so we are covered”
Spreading a service across zones is great for surviving a rack, a power feed or a single building. It is not a plan for losing a whole region, and it does nothing if the failure sits in the provider’s regional control plane, the part you need to launch replacements. The Atlas puts it simply: zone redundancy is a useful ingredient, yet it is not the same as recovery. A 2023 incident it recounts, where one of three buildings in a European cloud region went down, shows how awkward partial losses can get even for the provider’s own internal systems.
What to do instead: decide which kind of event you are planning for. A single-site loss may need nothing more than backups and zone spread. Regional events call for something that lives in another region.
Myth 2: “The hosting company handles disasters for us”
Hosts keep their platforms running, but they do not know which of your services matter most, what order to restore them in, or when to push the button. One large vendor’s own documentation, quoted in the Atlas, tells customers that DR does not come built in with the platform. Your contract buys capacity and reliability targets, not a rehearsed plan for your business.
What to do instead: write down who in your company owns each decision, and keep an account and permissions arrangement that still works when the primary region is struggling.
Myth 3: “Replication means we already have disaster recovery”
Replication copies changes quickly. That is precisely the problem when the change is a mistake. A buggy deployment, a bad migration or a corrupted table is mirrored across to the standby in seconds. The Atlas uses a July 2024 endpoint-software incident, in which a faulty content update crashed Windows machines worldwide, to illustrate that failing over does not help when the fault itself is everywhere.
What to do instead: keep point-in-time copies that you can roll back to, in addition to live replicas. Also remember that the data-loss window you actually get is set by the laggiest copy of anything the service depends on, including queues and search indexes, not only the main database.
Before the next one, a quick aside for anyone new to this: none of these corrections require special software. They need a calendar, a stopwatch and a willingness to look at awkward numbers.
Myth 4: “RTO is just our average repair time”
An objective is a promise about the worst case you will accept. Average repair time is a statistic about past incidents, and averages hide the one terrible day. Mixing them up leads to targets that feel safe on paper and collapse on the day the faulty change lands in front of a tired on-call engineer.
The Atlas also reminds readers that the clock runs from the beginning of the outage, so deciding to act counts. In its sample drill, thirty-one minutes of detecting and deciding plus thirty-three of technical work gave sixty-four minutes against a one-hour target. Nothing about the machines was slow; the people part tipped it over.
What to do instead: measure each drill in segments, and compute a latest-start time by taking the target, deducting the slowest recent technical time and a small cushion.
Myth 5: “RPO is how much data we can lose”
It sounds harmless, but the standards language is more precise. In NIST guidance, a recovery point objective names a point in time to which data must be restorable. The amount of data lost is a consequence of where that point falls, and busy hours can mean far more records lost than quiet ones for the same time gap.
What to do instead: state RPO as a time, and ask the business what a gap of that length would cost at peak. If the answer is alarming, tighten the objective or change the standby style.
Myth 6: “Once we have the root cause, we are back”
Understanding a failure and fixing it are different stages. In the Atlas timeline of a large 2025 US-region cloud event, the underlying DNS problem was identified early on, yet normal instance launches did not return until more than fourteen hours had passed. Anyone who equated diagnosis with restoration had no plan for that long tail, and spent the afternoon refreshing a status page. Treat the moment the cause is named as the start of the next phase, not the end of the incident.
What to do instead: track service restoration separately from cause analysis, and plan for the stage in which capacity is available again but queues, caches and dependencies are still catching up.
The pattern behind the myths
Notice how friendly each myth sounds. They persist because each carries a kernel of accuracy: zones do help, providers do work hard, replication is valuable, averages are worth tracking. Trouble begins once a partial truth gets promoted into a full plan, and the plan is never timed. A useful habit: put to any claim the question, “what event would make this stop being true?” and then see whether the answer appears in your drill scenarios.
Every myth above assumes that a feature already in place has removed the need to practise. The corrective is the same each time: rehearse with a stopwatch, write the numbers down, and let the results overrule the architecture diagram. Teams that work this way build confidence from evidence instead of from slides.
If you are building web systems for customers, these questions belong early in the design. Our engineering blueprint for web project requirements in depth and the question list in the questions to settle before commissioning a bespoke web app both leave room for recovery targets, and adding a line for them costs almost nothing at that stage.
For the reasoning, sources and further cases behind each correction, consult the DisasterRecoveryOps page in the Ops Atlas. Then pick the myth you most suspect your own team believes and test it with a drill this month.

