Safety management in aviation is built on a handful of unglamorous habits: report what you see, define who is accountable, rehearse the bad day and measure whether the controls work. Teams running software on public cloud need the same habits, and the practice that gathers them under one roof is CloudOps. What follows is eight of them, laid out as a plain checklist for people who would rather act than read theory. Each item says what to do, why it matters and how to tell it is working.
The source for the cloud side is the Ops Atlas from xOps School. Its page, the Atlas write-up on cloud operations, describes a young field with no neutral standard, so every item below is a suggestion to adapt, not a regulation to obey.
1. Draw the line between provider duties and yours
Start with a two-column sheet. On one side, list what the provider keeps running for you: the facilities, the physical network, the managed services themselves. On the other, list what remains with your team: who may log in, how services are configured, which data is backed up. Under the shared-responsibility model, a service being managed does not mean it is configured safely.
Proof it works: a new colleague can answer “whose job is this?” for any service in under a minute.
2. Give every account an owner and a purpose
An account with no owner is an uncontrolled stand-in for an unattended aircraft. Keep a register with the name of the responsible person, the business purpose and the date of the last review. Retire anything that cannot be explained.
Proof it works: every account on the bill appears in the register, and nothing appears in the register that is not on the bill.
3. Start new accounts from a template
Rather than configuring each account by hand, use a vending process so every new one arrives with the same logging, labels and access rules. The Atlas treats this as the heart of the foundation stage. Do not forget the trade-off it names: more accounts mean smaller areas affected by a failure and separate ceilings per account, but also more baselines to keep in line.
Proof it works: the share of accounts created through the template. The Atlas’s governed stage looks for around 90% of accounts to be vended and 95% to carry the baseline.
4. Report before you block
When you introduce a new rule, run it first in observation mode. Collect the events that would have been stopped, sort them into mistakes, genuine needs and clean-up jobs, and build the exception process from the second pile. Then switch to enforcement, organisational unit by organisational unit.
Proof it works: the first week of enforcement produces fewer surprises than the audit report predicted.
5. Know what each rule cannot see
Every restriction has gaps. The Atlas’s Region rules are a good example: they may not help with resources that already exist, they can stumble on globally scoped services, one provider’s location policy leaves out resource groups, and another’s constraint covers creation but not later data handling. Labelling rules have gaps too: on one platform, an unlabelled resource may not be flagged at all, and on another, labels are not inherited from the group above.
Write the gaps in the rule’s own documentation. A blind-spot list is a safety record in miniature.
Proof it works: an auditor can read your rule and learn what it does not cover without asking you.
6. Forecast ceilings before demand arrives
Every account has limits. Track usage, growth and the warning line, and calculate how many days remain. The Atlas’s worked example, a 256-unit limit with an 80% warning, 184 units in use and 2 units of daily growth, gives about 10.4 days until the warning fires and 36 days until exhaustion. Check the traps too: idle reservations count against the allowance, some GPU families start at zero and a provider can move the numbers itself.
Proof it works: zero launches fail because of a ceiling, which is the Atlas’s typical target.
7. Route provider notices to a person, not a void
Providers announce maintenance, retirements and incidents. Send those messages to a monitored shared address with a named duty owner, and log how fast each one is closed. In a 2025 global incident described by the Atlas, public status updates lagged by around an hour, so do not rely on a single channel; combine the status page with your own checks, hosted somewhere that does not share the same fate as the system being watched.
Proof it works: each notice has an owner, a due date and a closing note.
8. Rehearse the recovery, and count the management steps
A recovery plan that has never been tried is a hope. During a game day, simulate losing the primary location and count how many steps in the critical path depend on the provider’s management layer, such as creating resources or changing policies. As the Atlas puts it, a tighter recovery target should shrink that count, and the highest-priority failover path should show none.
Proof it works: the rehearsal result is written down, with the number of management-layer steps and the fixes assigned.
Bonus: keep any automated helper on a short leash
If you let an automated agent make changes in your cloud, the Atlas recommends that its write access stays in a setting where a human confirms every single change before it lands. It is the same principle as a second pair of eyes on a maintenance sign-off.
Answering the usual objections
“We are too small for this.” The habits scale down. A three-person team can keep an account register in a shared spreadsheet and a calendar reminder for quota reviews; the value comes from the routine, not the tooling.
“Rules will slow the developers.” Rules without an escape route will, which is why item 4 builds the exception path first. A request that takes a day to approve, with an expiry date attached, rarely causes resentment. A request that disappears into a queue causes workarounds.
“A second provider would solve most of this.” The Atlas disagrees. A second provider brings its own accounts, identity arrangements and rules to manage, so unless a named scenario demands it, a rehearsed plan inside one provider usually buys more resilience per hour of effort.
“Our provider guarantees high availability.” Read the fine print. The Atlas points out that a regional compute commitment of 99.99% pays out as bill credits when missed, and credits are not the same as a working service.
Measuring the whole list
The eight items map to a short scorecard: provisioning time, self-service rate, baseline coverage, Region conformance, tag compliance weighted by cost, quota runway, notice closure and management steps in failover. Pick three to start and review them monthly. Resist the urge to track everything at once; a scorecard nobody reads is the equivalent of an unread logbook.
The Atlas offers typical thresholds as a starting point: at least 95% for tag compliance and baseline coverage and none for quota-caused launch failures. Adjust them to your own risk and size, and write down the reason for any change.
Keeping the list alive
A checklist is only useful while it stays current. Put a recurring review on the calendar, quarterly for most teams, and use it to ask three things: which item produced evidence this quarter, which item has quietly lapsed and which provider change affects the wording of a rule. Assign each item a named owner so lapses have a face. When something lapses, record why. Often the reason, such as a reorganisation or a tool change, is more informative than the lapse itself, and it tells you which habit needs redesigning rather than reviving.
A realistic order of work
- Items 1, 2 and 7 first: they are cheap and reduce the chance of silent surprises.
- Items 3, 4 and 5 next: they need some design effort, but they make the estate consistent.
- Items 6 and 8 last: they need data and time, and they prove that the rest is working.
Nothing here requires a purchase. It requires attention, a shared register and a calendar. Teams who treat these habits as routine maintenance, and not as a special project, are usually the ones whose bad days turn out manageable.
