SAP-C03 · Domain 5: Operational Excellence and Automation · 20% of the exam
Task 5.3: Design observability and monitoring solutions.
Seeing a system: metrics, logs and traces with CloudWatch, X-Ray and OpenTelemetry, container monitoring, logs gathered from many accounts, alerting, AI/ML workload metrics, synthetic and real-user monitoring, and service-level objectives.
Study it
Metrics, logs and traces with CloudWatch, X-Ray and OpenTelemetry
Try each one before opening the answer. Every option is explained, with the AWS documentation page that proves it.
Question 1 · choose 1
A company runs microservices in 50 accounts. During incidents, engineers sign in to several accounts one after another to read metrics, logs and traces, which slows down diagnosis. The company wants one account in which engineers can search and correlate this telemetry from all 50 accounts without copying the data. What should a solutions architect do?
ACreate CloudWatch dashboards in each of the 50 accounts and share them publicly with the operations team
BUse CloudWatch cross-account observability with a monitoring account and the 50 accounts linked as sources
CCreate an organization trail in CloudTrail and search its log files during each incident
DExport all log groups from the 50 accounts to one S3 bucket every hour and query them with Amazon Athena
Show the answer and why
ACreate CloudWatch dashboards in each of the 50 accounts and share them publicly with the operations team
Incorrect
Sharing dashboards lets people view them, but engineers would still move between 50 separate dashboards and could not query across accounts.
BUse CloudWatch cross-account observability with a monitoring account and the 50 accounts linked as sources
Correct
With cross-account observability, a central monitoring account can view and interact with the metrics, logs and traces generated in the source accounts linked to it, and search across them.
CCreate an organization trail in CloudTrail and search its log files during each incident
Incorrect
CloudTrail records API activity in the accounts. It does not hold application metrics, logs or traces.
DExport all log groups from the 50 accounts to one S3 bucket every hour and query them with Amazon Athena
Incorrect
Exports copy the data and arrive late, they do not include metrics or traces, and they are a poor fit for live incident diagnosis.
"One place, all accounts, no copies" is CloudWatch cross-account observability with one monitoring account.
A new contract says that 99.9% of checkout API requests must complete in under 300 ms each month. Today the team alarms on the average latency reported by the Application Load Balancer, and the alarm never fires even though customers complain about slow checkouts. Which change makes the monitoring match the contract?
AAlarm on the Maximum statistic of TargetResponseTime above 300 ms
BAlarm on the p99.9 percentile of TargetResponseTime above 300 ms
CAlarm on the SampleCount statistic of TargetResponseTime above a fixed number
DLower the Average threshold from 300 ms to 100 ms so that the alarm fires earlier
Show the answer and why
AAlarm on the Maximum statistic of TargetResponseTime above 300 ms
Incorrect
Maximum reacts to a single slow request, so it would fire even when 99.9% of requests are fast and the contract is met.
BAlarm on the p99.9 percentile of TargetResponseTime above 300 ms
Correct
Percentile statistics show the value below which a given share of data points fall, so p99.9 measures the contract directly; averages hide slow requests.
CAlarm on the SampleCount statistic of TargetResponseTime above a fixed number
Incorrect
SampleCount counts requests. It says nothing about how long they took.
DLower the Average threshold from 300 ms to 100 ms so that the alarm fires earlier
Incorrect
A lower average threshold still hides the slowest requests and is not the measure the contract uses.
Translate the SLA into the same statistic: a percentile target needs a percentile alarm.
During an incident, engineers need to search the last two hours of JSON application logs in CloudWatch Logs for the slowest requests and group them by endpoint, without exporting the logs anywhere. Which tool fits?
AA metric filter created after the incident
BAn export of the log group to S3 followed by Athena queries
CCloudWatch Logs Insights queries on the log group
DAWS Config advanced queries
Show the answer and why
AA metric filter created after the incident
Incorrect
Metric filters count matches in new log data; they do not search past logs by endpoint.
BAn export of the log group to S3 followed by Athena queries
Incorrect
Exporting adds delay and is what the engineers want to avoid.
CCloudWatch Logs Insights queries on the log group
Correct
Logs Insights lets you interactively search and analyze log data in CloudWatch Logs with queries.
DAWS Config advanced queries
Incorrect
Config queries resource configuration, not application logs.
Ad hoc log analysis in place is CloudWatch Logs Insights.
When the checkout service degrades, its CPU, latency, 5xx and health check alarms each page the on-call engineer, often a dozen times per incident. The team wants a single page only when the 5xx alarm and at least one of the latency or health check alarms are in ALARM at the same time. No page may go out while a deployment alarm signals a planned release, the individual alarms must stay in place for diagnosis, and the team will not write code for this. Which approach meets these requirements?
ACreate an EventBridge rule that matches state changes of the four alarms and sends each matching event to the pager integration
BRaise the evaluation periods and datapoints to alarm on each of the four alarms so that they fire less often
CPage from a composite alarm whose rule requires the 5xx alarm and either of the other two, with the deployment alarm as suppressor
DSend all four alarms to one SNS topic and give the pager subscription a filter policy that passes only messages from the 5xx alarm
Show the answer and why
ACreate an EventBridge rule that matches state changes of the four alarms and sends each matching event to the pager integration
Incorrect
CloudWatch sends an event to EventBridge whenever an alarm changes state, so a rule can route those events to a pager. Each matching event is still a separate page, and combining alarm states would need code.
BRaise the evaluation periods and datapoints to alarm on each of the four alarms so that they fire less often
Incorrect
Datapoints to alarm sets how many breaching data points within the evaluation periods put an alarm into ALARM, which filters out short spikes. During a real incident, every alarm still pages on its own.
CPage from a composite alarm whose rule requires the 5xx alarm and either of the other two, with the deployment alarm as suppressor
Correct
Composite alarms combine the states of other alarms with Boolean rule expressions and act only at the aggregated level, which reduces alarm noise. A suppressor alarm stops the composite alarm's actions during planned deployments, and the underlying alarms stay in place for diagnosis.
DSend all four alarms to one SNS topic and give the pager subscription a filter policy that passes only messages from the 5xx alarm
Incorrect
A subscription filter policy delivers only the subset of messages that match it. Filtering on one alarm pages on 5xx errors alone, ignores the required combination and still pages during releases.
The constraints are one page for a specific combination of alarms, silence during planned releases, the original alarms kept, and no code. EventBridge routing and SNS filtering each pass single alarm notifications, and tuning datapoints only makes each alarm less sensitive. A composite alarm with a rule expression and a suppressor alarm meets all of them.
A multi-tenant API serves about 4,000 tenants and writes JSON access logs with tenantId, route and status fields to CloudWatch Logs. Twice last month one tenant's faulty integration flooded the API with failing requests and slowed it for everyone, and finding that tenant took an hour. Operations wants a continuously updated ranking of the tenants that cause the most 5xx responses, graphed over time, and an alarm when any single tenant exceeds 500 errors per minute. The team does not want a custom metric for each tenant and cannot change the application. Which solution meets these requirements?
ACreate a Contributor Insights rule on the log group for 5xx responses keyed on tenantId, and alarm on its top contributor value
BCreate an anomaly detection alarm on the API's total 5xx metric so that unusual error levels page the on-call team
CAdd a metric filter for 5xx responses with tenantId as a dimension, and alarm on the metric that the filter publishes
DSave a CloudWatch Logs Insights query that groups 5xx responses by tenantId, and run it when the API slows down
Show the answer and why
ACreate a Contributor Insights rule on the log group for 5xx responses keyed on tenantId, and alarm on its top contributor value
Correct
Contributor Insights analyzes high-cardinality log data into time series of the top-N contributors. The INSIGHT_RULE_METRIC function with MaxContributorValue returns the top contributor's count for each period, and an alarm can be set on it.
BCreate an anomaly detection alarm on the API's total 5xx metric so that unusual error levels page the on-call team
Incorrect
Anomaly detection compares a metric with a band of expected values. On the total error metric it can page during a flood, but it cannot rank tenants or say which tenant is responsible.
CAdd a metric filter for 5xx responses with tenantId as a dimension, and alarm on the metric that the filter publishes
Incorrect
Each tenantId value becomes its own custom metric, which the team wants to avoid. AWS warns against high-cardinality dimensions and may disable a metric filter that generates 1,000 different dimension values.
DSave a CloudWatch Logs Insights query that groups 5xx responses by tenantId, and run it when the API slows down
Incorrect
Logs Insights is for interactive searches during an investigation. A query that someone runs by hand is not a continuously updated ranking, and it pages nobody when a tenant crosses the limit.
The requirement is a top-talkers view over thousands of keys, which is what Contributor Insights is for: one rule on the existing JSON logs ranks tenants continuously, and an alarm on MaxContributorValue fires when the worst tenant crosses 500 errors per minute, all without per-tenant metrics or code changes. Metric filters with a tenant dimension create the per-tenant metrics the team wants to avoid.
Server metrics for a web application look healthy, but customers complain that pages load slowly in some countries and browsers. The team wants performance data from actual user sessions. Which tool fits?
AAWS Trusted Advisor performance checks
BCloudWatch RUM for the web application
CVPC Flow Logs for the web subnets
DCloudWatch metrics for the load balancer
Show the answer and why
AAWS Trusted Advisor performance checks
Incorrect
Trusted Advisor checks accounts, not user page loads.
BCloudWatch RUM for the web application
Correct
CloudWatch RUM collects client-side performance data from actual user sessions in near real time.
CVPC Flow Logs for the web subnets
Incorrect
Flow logs record IP traffic, not page load times.
DCloudWatch metrics for the load balancer
Incorrect
Load balancer metrics are measured on the server side, not in users' browsers.
Real user performance is measured with CloudWatch RUM.
Workers that process an SQS queue sometimes hang without crashing, and orders silently wait for hours before anyone notices. Which alarm would detect this condition quickly?
AAn alarm on the queue's message size
BAn alarm on the CPU of the worker instances
CAn alarm on the number of messages sent to the queue
DAn alarm on the queue's ApproximateAgeOfOldestMessage
Show the answer and why
AAn alarm on the queue's message size
Incorrect
Message size does not change when consumers hang.
BAn alarm on the CPU of the worker instances
Incorrect
Hung workers can show normal or low CPU.
CAn alarm on the number of messages sent to the queue
Incorrect
Producers keep sending normally while consumers hang.
DAn alarm on the queue's ApproximateAgeOfOldestMessage
Correct
This metric reports the age of the oldest unprocessed message, which grows when consumers stop processing.
Stalled consumers show up as a growing age of the oldest message.
A company migrating Kubernetes workloads to EKS has dashboards and alerts built on Prometheus queries. It wants to keep PromQL but not run its own Prometheus servers at scale. Which target fits?
AAWS X-Ray traces
BCloudWatch Logs Insights only
CSelf-managed Prometheus on larger EC2 instances
DAmazon Managed Service for Prometheus
Show the answer and why
AAWS X-Ray traces
Incorrect
Traces follow requests; they do not replace Prometheus metrics.
BCloudWatch Logs Insights only
Incorrect
Logs Insights queries logs, not Prometheus metrics with PromQL.
CSelf-managed Prometheus on larger EC2 instances
Incorrect
The company would still run Prometheus servers.
DAmazon Managed Service for Prometheus
Correct
It is a serverless, Prometheus-compatible monitoring service for container metrics.
Prometheus-compatible metrics without servers is Amazon Managed Service for Prometheus.