cert-bank

Guides

이 글을 한국어로 읽기

SAP-C03 Common Mistakes: Four Families of Wrong Answers on the Solutions Architect – Professional Exam

Wrong answers on the AWS Certified Solutions Architect – Professional exam (SAP-C03) are rarely silly. They are usually good designs that solve a slightly different problem from the one in the question. Reading back through the explanations for all 509 SAP-C03 questions in our practice bank, the tempting wrong options keep falling into four families. Each family has a test you can apply before you choose, and knowing which family you tend to fall into is worth more than another hundred questions answered blind.

Study guide · Common mistakes · AWS Certified Solutions Architect – Professional (SAP-C03)

SAP-C03 Common Mistakes: Four Families of Wrong Answers on the Solutions Architect – Professional Exam

Why SAP-C03 punishes "almost right"

SAP-C03 replaces SAP-C02 from 17 November 2026: 75 questions (65 scored, 10 unscored), passing score 720. The questions are long (about 45 words on average in our bank, against about 26 for the AI Business Strategist exam) and 64% are use-case scenarios. Each scenario carries several requirements at once, such as a recovery objective, a compliance rule, a cost limit and "least operational overhead", and three of the four options usually meet most of them. The exam is decided by the one requirement an option misses.

Family 1: fixing the wrong objective

The design addresses a real weakness, just not the one the question measured. Recovery questions are the classic case, because RTO (how long recovery takes) and RPO (how much data is lost) sound alike and have completely different fixes.

Scenario from our bankTempting answerWhat the explanation says
Pilot light recovery took 4 hours instead of 30 minutes because engineers launched servers and changed DNS by hand.Relax the RPO, or rely on backups instead.The time went into rebuilding the app tier. That is an RTO problem: prebuilt AMIs or IaC, automated runbooks and DNS failover, tested.
Warm standby failover lost 45 minutes of transactions against a 5-minute RPO.Faster Auto Scaling, shorter health checks, a bigger standby.Lost data equals the replication lag at the moment of failure. Use lower-lag replication (for example Aurora Global Database) and alarm on lag.
A generative AI latency SLO says 90% of calls per model finish in 4 seconds; the average looks healthy while users complain.Lower the threshold on the average.An average hides the slow tail. Alarm on the p90 statistic, per model, which is what the SLO actually states.

Test before choosing: name the number that failed (minutes of downtime, minutes of data, a percentile, a cost line) and reject every option that does not change that number.

Family 2: detecting when the requirement says prevent

Organisation-scale governance is a heavy theme in our SAP-C03 set: AWS Organizations is the most frequently tagged service there, well ahead of Amazon S3 (61 questions against 44). The trap is choosing a control that reports a violation when the question demands that the violation never happens.

  • Data residency across the organisation. An AWS Config rule flags buckets created in the wrong Region; an EventBridge rule could delete them. Both act after the data has landed. The answer that prevents creation is an SCP on the OU that denies the actions when aws:RequestedRegion is not approved.
  • Mandatory tags for chargeback. Cost allocation tags and Cost Explorer reports only show what is missing. An SCP makes the tag required at creation; a tag policy keeps its values standard.
  • Guardrails for a public chatbot. Pointing production at a guardrail's working draft means every edit, tested or not, goes live at once, and nobody can show which policy was in force on a given date. Test the draft, publish a numbered version, and move production to it after approval.

The words to watch for are "prevent", "ensure that no", "must never" and "before". If they appear, Config rules, dashboards and alerts are supporting actors at best.

Going deeper: why an SCP holds where a permissions boundary does not

A permissions boundary is attached to one user or role inside an account, so whoever administers that account can create a new role without it or remove it. An SCP is attached centrally in AWS Organizations to the organisation root, an OU or an account, and it limits every IAM user and role in the affected member accounts, including each account's root user (service-linked roles are the exception). A member account's administrator cannot detach it or grant around it, because an SCP never grants anything: it sets the maximum, and an explicit deny in it ends the evaluation whatever the account's own policies say. Two limits are worth remembering for the exam. SCPs do not apply to the management account, which is one reason workloads should not run there. And a Region-deny SCP needs an exemption (NotAction) for global services such as IAM, CloudFront and Route 53, whose requests go to a single endpoint in us-east-1 and would otherwise be denied.

Family 3: paying for capacity that sits idle

Cost-Optimized Architecture Design is 15% of the exam, and its wrong answers are rarely wasteful in an obvious way. They buy a commitment or a fixed fleet where the scenario describes variable use.

  • Reserved Instances bought for one AZ when the workload has moved to an Auto Scaling group across three. Buying more Zonal RIs adds commitment without fixing the mismatch; changing the scope to Regional lets the discount follow the instances.
  • Non-production environments used part of the day. Three-year Reserved Instances or bigger instances still pay for the idle hours. Schedule them off.
  • Interruption-tolerant, checkpointed batch jobs. Diversified Spot capacity, not reservations or dedicated hardware.
  • Predictable day and night traffic. Scheduled scaling plus target tracking, not a fleet sized for the peak.

The pattern behind all four: commit only for the steady baseline, and let the variable part follow demand.

Going deeper: what a Reserved Instance actually is

A Reserved Instance is not a server. It is a billing discount that AWS applies, hour by hour, to running instances that match its attributes. A zonal RI matches one instance type in one Availability Zone and, in exchange, also reserves capacity there. A regional RI gives up the capacity reservation but matches usage in any AZ of the Region and, for Linux instances with default tenancy, across sizes in the same family, using a normalisation factor. When an Auto Scaling group moves instances between zones, a zonal RI simply finds nothing to match and the discount goes unused. Standard RIs can have their scope modified, which is why the bank's answer changes the scope rather than buying more or giving up on the commitment.

Family 4: the name fits, the mechanism does not

These options mention the right family of service but miss a property the scenario requires. They are the hardest to spot because the keyword match is so strong.

RequirementNear miss → what meets it
Keys must stay in single-tenant hardware under the company's exclusive control; the database uses PKCS#11KMS (multi-tenant, and no PKCS#11 interface) → AWS CloudHSM used directly
Every Region must accept writesAurora Global Database (one writer Region) → DynamoDB global tables
Tenants in different jurisdictions with different residency rulesGlobal tables (copy every item everywhere) → one cell per jurisdiction, tenants routed to their home cell
Employees need single sign-on to many accountsAmazon Cognito (application users) → IAM Identity Center permission sets
Static IP addresses for allow-lists plus fast Regional failoverRoute 53 (changes addresses through DNS) or CloudFront → AWS Global Accelerator
A RAG assistant misses exact part numbers and ranks the right passage lowChanging the temperature or using bigger chunks → hybrid search plus a reranker

Test before choosing: for the option you like, say out loud the one property the scenario insists on ("single-tenant", "every Region writes", "static IP") and check that the service has that property, not just a related name.

Beginner tip: for your next two practice sets, log every miss in a table with three columns: the requirement word you missed (for example "prevent" or "5-minute RPO"), the option you chose, and the family it belongs to (1 to 4). After about twenty misses, one or two families will hold most of them. Spend your remaining study time there instead of on new topics.

SAP-C03: four tests to run on the option you like
Does it leave unchanged the number that failed (downtime, data lost, a percentile, a cost)?
YES →
Family 1: wrong objective. Reject
↓ no
Does the stem say prevent or must never, while the option only reports?
YES →
Family 2: detects, not prevents. Reject
↓ no
Does it buy a commitment or a fixed fleet for variable use?
YES →
Family 3: idle capacity. Reject
↓ no
Does the service lack the one property the scenario insists on?
YES →
Family 4: name fits, mechanism does not. Reject
↓ no
OTHERWISE →
The option survives all four tests: choose it
Run the tests in order on your chosen option; the first yes names its family and rules it out.

Where the new SAP-C03 content fits

SAP-C03 adds generative AI and agents (Amazon Bedrock is a tagged service in 28 of our 509 questions), DevSecOps, chaos engineering and post-quantum cryptography. None of this needs new instincts. A guardrail on its working draft is a Family 2 mistake; tuning the model when retrieval is broken is Family 4; a chaos experiment watched by an engineer on a dashboard, instead of a stop condition on a customer-facing alarm that halts it automatically, is Family 2 again. Learn the new services, but apply the same four tests.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Exam facts for SAP-C03 as announced by AWS at the time of writing; scenarios and counts come from our own practice bank. Check the official exam guide for the current domains and format.