Professional Cloud Architect · Domain 6: Ensuring solution and operations excellence · 12% of the exam
Task 6.5: Evaluating quality control measures
Measuring quality with SLOs and error budgets, release gates, vulnerability scanning and checks on infrastructure changes.
Study it
Quality control: error budgets, release gates and scanning
Lesson coming
Sample questions
Try each one before opening the answer. Every option is explained, with the Google Cloud documentation page that proves it.
Question 1 · choose 1
A service has an availability SLO of 99.9% over 28 days. Halfway through the period its error budget is already spent, and product managers want to ship two new features this week. What should the team do under an error budget policy?
ALower the SLO to 99% for the rest of the period so the releases can go ahead
BShip the features and disable the SLO alerting policies until the features are stable
CPause risky releases and put reliability work first until the budget recovers
DShip the features now and recalculate the error budget at the next period
Show the answer and why
ALower the SLO to 99% for the rest of the period so the releases can go ahead
Incorrect
Moving the target to fit the releases removes the SLO's value as a quality control.
BShip the features and disable the SLO alerting policies until the features are stable
Incorrect
Silencing alerts hides the problem while users keep getting a degraded service.
CPause risky releases and put reliability work first until the budget recovers
Correct
A depleted error budget means further risky changes such as new releases are likely to break the SLO, so the policy shifts effort to reliability.
DShip the features now and recalculate the error budget at the next period
Incorrect
The budget is already spent for this period; shipping now knowingly breaks the objective.
Error budgets turn an SLO into a release gate: while budget remains, the team can take risk; once it is gone, reliability work comes first.
A team is defining service level indicators for a public REST API to measure its quality. Which TWO measurements make good SLIs? (Choose TWO.)
AThe average CPU utilization of the API's backend nodes
BThe number of Pods running for the API in the cluster
CThe proportion of requests served without a 5xx error
DThe number of deployments the team ships each week
EThe proportion of requests served in under 300 milliseconds
Show the answer and why
AThe average CPU utilization of the API's backend nodes
Incorrect
CPU use is a resource metric; users can see slow or failed requests while CPU looks fine.
BThe number of Pods running for the API in the cluster
Incorrect
Replica counts describe capacity, not the quality users receive.
CThe proportion of requests served without a 5xx error
Correct
Availability measured as the share of good requests reflects what users experience.
DThe number of deployments the team ships each week
Incorrect
How often the team ships says nothing about the quality users receive.
EThe proportion of requests served in under 300 milliseconds
Correct
Latency measured as the share of fast enough requests is a standard user-facing SLI.
Quality control starts with measuring what users experience. SLIs such as the ratio of good to total requests for availability and latency feed SLOs and error budgets.
A team changes the prompts and model version of a customer-facing Gemini assistant every few weeks. Some changes have made answers worse in ways nobody noticed until customers complained. What quality control should the architect add before each change ships?
ARun the Gen AI evaluation service on a fixed test set and compare the scores
BLower the model's temperature so that its answers are more consistent
CMonitor the endpoint's latency and error rate after each change has shipped to users
DAsk one developer to try a few questions in the console before release
Show the answer and why
ARun the Gen AI evaluation service on a fixed test set and compare the scores
Correct
The evaluation service scores model outputs against criteria, so a candidate can be compared with the current version before release.
BLower the model's temperature so that its answers are more consistent
Incorrect
Lower temperature changes randomness, not whether the new prompt is better.
CMonitor the endpoint's latency and error rate after each change has shipped to users
Incorrect
Operational metrics do not catch answers that are wrong but delivered quickly.
DAsk one developer to try a few questions in the console before release
Incorrect
A handful of manual checks is not a repeatable quality gate.
Generative AI changes need their own quality gates. A fixed evaluation set scored by the evaluation service turns subjective quality into a release decision.
A data pipeline loads sales tables in BigQuery that feed executive dashboards. Recently, null customer IDs and duplicated orders reached the dashboards. The team wants rules such as non-null and uniqueness checked on every load, with alerts when they fail. What should the architect use?
ABigQuery table snapshots taken after each load
BCloud Monitoring alerts on the runtime of each of the BigQuery load jobs
CColumn-level access control on the customer column
DKnowledge Catalog auto data quality scans with rules on the tables
Show the answer and why
ABigQuery table snapshots taken after each load
Incorrect
Snapshots preserve data for recovery; they do not check its quality.
BCloud Monitoring alerts on the runtime of each of the BigQuery load jobs
Incorrect
Job duration says nothing about nulls or duplicates in the results.
CColumn-level access control on the customer column
Incorrect
Access control limits who sees a column, not what values it holds.
DKnowledge Catalog auto data quality scans with rules on the tables
Correct
Auto data quality validates BigQuery data against defined rules on a schedule and can log alerts when the data does not meet them.
Quality control applies to data products as well as code. Rule-based data quality scans catch bad loads before they reach decision makers.
A platform team wants Kubernetes manifests to follow rules, such as no privileged containers and required cost-center labels. It wants violations blocked at admission in the GKE clusters and also caught earlier, before changes are merged. Which TWO measures should the team use? (Choose TWO.)
AKubernetes RBAC roles that limit who can deploy
BPolicy Controller constraints enforced in the clusters
CThe same constraints checked against manifests in CI
DBinary Authorization attestations on every image
ENetworkPolicies that isolate each team's namespace
Show the answer and why
AKubernetes RBAC roles that limit who can deploy
Incorrect
RBAC controls who may act, not whether a manifest follows the rules.
BPolicy Controller constraints enforced in the clusters
Correct
Policy Controller enforces and audits programmable policies in clusters as guardrails for security and compliance.
CThe same constraints checked against manifests in CI
Correct
Policy Controller supports a shift-left approach, catching non-compliant changes before they reach the cluster.
DBinary Authorization attestations on every image
Incorrect
Binary Authorization checks images, not the settings or labels in manifests.
ENetworkPolicies that isolate each team's namespace
Incorrect
NetworkPolicies control Pod traffic, not manifest content.
Quality control for infrastructure works best in layers: the same policies checked early in CI and enforced at admission time.