Skip to content
On this page 7

What production-ready actually means

Production-ready is not a feeling. It is eight properties, each with a question you can answer about your own estate this week, with evidence.

Production-ready is not a stage a system reaches. It is eight properties a system either holds or does not: secure, reliable, observable, scalable, automated, recoverable, compliant, operable. Each property asks one question, and the only answer that counts is evidence produced recently. Everything else is an opinion held by someone who has not been paged yet.

This post turns that into a tool. By the end you will have eight questions, a scorecard you can draw on a whiteboard, and a way of running the evaluation on your own estate in about ninety minutes — with no purchase involved and no consultant in the room.

Why “it works” is not an answer

#

The most expensive sentence in infrastructure is “it works.” It is true and useless at the same time, because it describes the system on a day when nothing was wrong.

Production is not the state of working. Production is the state of continuing to work while a dependency is failing, while the person who built it is asleep, while an auditor is asking for records and while traffic is four times what anyone planned for. A system is production-ready when those four sentences have boring answers.

That is why the properties are phrased as questions with evidence attached. “Is it secure?” invites a reassuring answer. “Show me who had production write access last quarter, and who removed the two who left” does not.

The eight properties, and the question each one asks

#

This is the BuiltForProd Standard used as an evaluation framework rather than a marketing list. Read the middle column as the question you ask out loud, and the right-hand column as what a good answer looks like.

PropertyThe question it asksEvidence that answers it
SecureWho can change production, and how is that enforced rather than agreed?An access list from the identity system, a preventive policy that denies the rest, and no long-lived credentials anywhere
ReliableWhat happens when one component, one zone or one dependency fails?A dependency map, a recent failure that was absorbed, and alerting that fired before a customer noticed
ObservableHow long between a symptom and knowing which component caused it?A trace or a dashboard that answered a real question last month, and logs from every environment in one place
ScalableWhat breaks first under four times the load, and do you know the number?A load test with a named bottleneck, autoscaling that has actually scaled, and a quota review
AutomatedCan the system be rebuilt without a specific person and a specific laptop?A pipeline run that created an environment end to end, and infrastructure defined as code with reviewed changes
RecoverableWhen did you last restore from a backup, and how long did it take?A dated restore drill with a measured recovery time, compared against the target you published
CompliantCan you produce the last quarter’s access records and change approvals today?Configuration history, audit logs in a separate account, and change approvals that are a by-product of doing the work
OperableCan the second-best engineer run this alone at 3 AM?A runbook someone other than the author used, an alert that means action, and an escalation path with names

Two details make this a tool rather than a checklist.

Each question has a date. Not “do we have backups” but “when did we last restore one”. A property is evidenced or it is assumed, and evidence expires. Ninety days is a reasonable window for most of these; a year is not.

Each question has an owner. The most common real finding is not that a property fails. It is that nobody is responsible for it, which means it will fail eventually and the failure will surprise everyone.

Figure 1
The eight properties passing through an evidence gate

Eight properties — secure, reliable, observable, scalable, automated, recoverable, compliant and operable — each enter a gate labeled: show me evidence from the last ninety days. Properties with evidence leave as verified. Properties without evidence leave as recorded assumptions carrying an owner and a review date.

securereliableobservablescalableautomatedrecoverablecompliantthe evidence gate”show me the lastninety days”verifieddated, ownedassumptionowner, review dateoperable is evaluated last, because it is the property every other one depends on
The evaluation is a gate, not a grade. Each property enters the same test — produce evidence from the last ninety days — and leaves as either a verified property or a recorded assumption with an owner and a date.

How to run this on a whiteboard this week

#

Ninety minutes, four people, one room. Invite the engineer who is on call, the engineer who wrote most of the infrastructure, someone from security or whoever answers customer questionnaires, and the person who owns the roadmap. That last one matters: the output is a set of trade-offs, and trade-offs need the person who decides them.

Draw five columns:

PropertyQuestionEvidenceOwnerLast verified
SecureWho can change production?Identity system export, preventive policy
ReliableWhat absorbs a single failure?Dependency map, recent absorbed failure
ObservableSymptom to cause, in minutes?The last incident timeline
ScalableWhat breaks first at 4x?Load test result, quota review
AutomatedRebuild without a person?Pipeline run that built an environment
RecoverableLast successful restore?Drill record with measured time
CompliantLast quarter’s access records?Configuration history, audit log query
OperableCan the second-best engineer run it?A runbook used by someone else

Then three rules that make the exercise honest:

  1. Nobody fills in “evidence” from memory. If it cannot be produced in the room or linked to, it goes in the assumption column. This is uncomfortable for about ten minutes and then becomes the most useful part of the meeting.
  2. An empty owner cell is a finding on its own, and usually the most important one.
  3. Do not total it. Write the assumptions in priority order instead.

Reading the result

#

Some patterns recur often enough to name.

Six green, two blank, and the blanks are recoverable and operable. The most common result in a fast-growing company, and the most dangerous, because both blanks only cost you during an incident. A backup job that has never been restored is a hypothesis — the argument is in backups you haven’t restored are assumptions — and a system only one person can run fails the 3 AM test.

Everything evidenced except compliant. Usually a young company that has not been asked yet. It is a timing problem rather than an engineering one: the observation window is the constraint, so switching on the recorders is the cheapest action available and it gets cheaper the earlier you do it. See SOC 2: what your infrastructure has to produce.

Secure is evidenced but automated is not. This one is a trap. Controls enforced by convention rather than by code survive exactly as long as the people who remember them. If the identity model lives in a console and not in a reviewed repository, the secure row is weaker than it looks.

Three or more rows with no evidence and no owner. At that point you do not have a set of gaps; you have a foundation that was never finished. Fixing the rows one at a time usually costs more than rebuilding the layer underneath them, which is the real argument for a foundation rather than a backlog. We put actual numbers on that trade-off in the true cost of building your own landing zone.

What each property costs to evidence

#

Rough guide to sequencing, cheapest first. All of these are things you can start without buying anything.

PropertyCheapest useful evidenceEffort
CompliantTurn on audit logging and configuration recording, todayHours
AutomatedRun the pipeline into an empty environment and watch it fail; fix until it does notDays
ObservableTime the last incident: page to cause, in minutesHours, plus honesty
SecureExport who has production access; remove the ones who leftHours
RecoverableRestore one production database into an isolated account and measure itOne to two days
OperableHand a runbook to an engineer who did not write it and watch, silentlyHalf a day
ReliableKill one instance in staging during business hoursHalf a day
ScalableLoad test to four times peak and name the first bottleneckTwo to three days

The ordering is deliberate: the top four produce evidence that is also useful for a customer security review, so the work pays twice.

The honest limits of this framework

#

It does not tell you whether your architecture is good. Two systems can both hold all eight properties and one of them can still be the wrong design for your business. The properties are about survivability and operability, not about fitness for purpose.

It also does not make you compliant with anything. The compliant property says your infrastructure produces evidence continuously. Certification is a licensed auditor’s opinion about your whole organization — people, policies, vendors — and no product, ours included, can deliver it.

And it is not a one-time exercise. Evidence expires. The point of the date column is that a property verified fourteen months ago is a property you are currently assuming.

Where to take it from here

#

Run the whiteboard version first. It costs ninety minutes and the argument itself is worth more than most audits.

If the result is three or more rows nobody owns, the next step is to decide whether you are fixing rows or replacing the layer underneath them. The BuiltForProd Assessment is the outside version of exactly this exercise, measured against the BuiltForProd Standard, and the BuiltForProd Baseline is what it looks like when the eight properties are built into the foundation rather than added to it — starting with the account boundaries described in the multi-account landing zone, explained.

Questions

What does production-ready mean?

Production-ready means a system holds eight properties — secure, reliable, observable, scalable, automated, recoverable, compliant and operable — and that each one can be demonstrated with recent evidence rather than asserted. The practical test is whether a competent engineer who did not build it would agree to be paged for it.

How do we test whether our own systems are production-ready?

Take the eight properties, write the question each one asks on a whiteboard, and for every system write the evidence, the owner and the date it was last verified. Any row where the evidence is a person saying it should be fine is an assumption. The exercise takes about ninety minutes and needs the four people who would actually be paged.

Is there a score for production readiness?

We do not publish a numeric score, deliberately. A single number lets a failing property hide behind seven passing ones, and it invites teams to optimize the number instead of the property. The useful output is a list of properties with no evidence and no owner, in priority order.

Which property should we fix first?

Usually recoverable, then operable. A restore you have never performed and a system only one person can run are the two failures that turn an incident into an outage of unpredictable length, and both are cheap to test compared with what they cost during the event.

Does being production-ready mean being compliant?

No. The compliant property means your infrastructure produces the evidence an auditor asks for — access records, configuration history, change approvals, log integrity — continuously. Certification is an opinion issued by a licensed auditor about your whole organization, and no infrastructure can grant it.