10 free sample questions from a bank of 10, with the correct answers and explanations. No signup required — start practising right now.
1A CloudOps engineer is examining the following AWS CloudFormation template:
Why will the stack creation fail?
The Outputs section of the CloudFormation template was omitted.
The Parameters section of the CloudFormation template was omitted.
The PrivateDnsName cannot be sot from a CloudFormation template.
The VPC was not specified in the CloudFormation template.
Answer: C
The short version
C — PrivateDnsName is not a settable property in a CloudFormation template, so the create fails.
CloudFormation rejects a resource that specifies a property the resource type does not support, which is why the stack never gets past validation.
Key concepts
CloudFormation resource property: only attributes defined by the resource type schema may appear in a resource block; an unsupported property fails template validation before any resource is created.
PrivateDnsName: a network interface attribute that assigns a hostname to an interface, but for EC2-based network interfaces the private DNS name is managed by the VPC DHCP option set, not set per resource.
DHCP options set: the VPC-level configuration that supplies domain-name and domain-name-servers, which is how instances receive their private DNS names.
Template validation: CloudFormation checks the template against resource schemas and reports errors such as unsupported or malformed properties at create time.
Elastic Load Balancing / Auto Scaling resource types: resource types such as AWS::ElasticLoadBalancingV2::TargetGroup and launch templates support very different property sets, which is where this kind of mistake usually appears.
Why C is correct
The template sets a private DNS name on a resource inside the stack, and CloudFormation's resource schema for that type has no PrivateDnsName property. Template validation therefore fails, the stack create is rejected, and no resources are created at all — which is exactly the "why will the stack creation fail" behaviour the question describes. The reason is structural, not a permissions or ordering issue: the value cannot be expressed in a CloudFormation resource at all, so it must be assigned outside the stack through the VPC's DHCP options or by configuring the hostname on the network interface after it exists.
Why the others are wrong
A. The Outputs section is entirely optional. CloudFormation creates stacks successfully with no outputs at all; outputs exist only to publish values for other templates or tooling to consume.
B. The Parameters section is also optional, and its absence cannot cause a failure. A parameter can only be an input, so with no parameters declared there is simply nothing to supply.
D. Not naming a VPC in a template is only a problem if the resource type genuinely requires a VPC property, and many resource types infer the VPC from a referenced subnet or security group. It is not a universal validation rule, so it is not the general cause of a failed create.
SOA-C03 exam tip
CloudFormation validates schema before it creates anything. If a question shows a template and asks why the create fails, look for an unsupported or invalid property value first — missing Parameters and Outputs are always optional.
2A company applies user-defined tags to resources that are associated with the company's AWS workloads. Twenty days after applying the tags, the company notices that it cannot use the tags to filter views in the AWS Cost Explorer console.
What is the reason for this issue?
It takes at least 30 days to be able to use tags to filter views in Cost Explorer.
The company has not activated the user-defined tags for cost allocation.
The company has not created an AWS Cost and Usage Report.
The company has not created a usage budget in AWS Budgets.
Answer: B
The short version
B — The tags exist on resources but have not been activated for cost allocation in AWS Cost Explorer.
Cost Explorer can only filter on tags that AWS has ingested and activated, which is a separate step from tagging resources.
Key concepts
User-defined cost allocation tags: key/value pairs applied to resources (or activated in specific services) and used to slice cost and usage data.
Tag activation: in AWS Cost Explorer you must explicitly activate a tag key before it becomes available for filtering; activation can also be driven by AWS Organizations tag policies.
Cost and Usage Report: the detailed, near-real-time billing dataset used for Athena or custom analysis; it is a different mechanism from Cost Explorer's tag filters.
Cost Explorer: the built-in cost and usage visualisation tool that supports filtering by activated tags, with a period granularity of daily or monthly.
AWS Budgets: an alerting and planning service that can be configured with tag-based filters, but it is not a prerequisite for Cost Explorer filtering.
Why C is correct
Tagging a resource only records the tag on that resource; it does not automatically publish the tag to the billing system. For Cost Explorer to filter on a key, that key must be activated in the Cost Explorer tag activation page (or be pushed by an Organizations tag policy), and AWS then ingests it on its normal schedule. Until activation happens, the console offers the tag as unavailable and filters return nothing, which is precisely the symptom described — twenty days is far longer than any propagation delay. The distinction to hold onto is tagged versus activated for cost allocation.
Why the others are wrong
A. There is no 30-day wait before tags can be used as Cost Explorer filters. Activation is an explicit user action, and once activated the key becomes selectable within about a day; waiting does not substitute for activating it.
C. A Cost and Usage Report is required if you want to query billing line items yourself with Athena or CloudFormation, but Cost Explorer has its own built-in cost data and filters. A missing CUR explains not having a report, not a tag filter being greyed out.
D. Budgets is an optional cost-management and alerting tool. You can use Cost Explorer tag filters with or without any budget, so the absence of a budget cannot block filtering.
SOA-C03 exam tip
Tag it, then activate it. In cost questions, "I can see the tag on the resource but cannot filter on it in Cost Explorer" always means tag activation for cost allocation, never a 30-day delay.
3An environment consists of 100 Amazon EC2 Windows instances. The Amazon CloudWatch agent is deployed and running on all EC2 instances with a baseline configuration file to capture log files. There is a new requirement to capture the DHCP log files that exist on 50 of the instances.
What is the MOST operationally efficient way to meet this new requirement?
Create an additional CloudWatch agent configuration filo to capture the DHCP logs. Use the AWS Systems Manager Run Command to restart the CloudWatch agent on each EC2 instance with the append-config option to apply the additional configuration file.
Log in to each EC2 instance with administrator rights. Create a PowerShell script to push the needed baseline log files and DHCP log files to CloudWatch.
Run the CloudWatch agent configuration file wizard on each EC2 instance. Verify that the baseline log files are included and add the DHCP log files during the wizard creation process.
Run the CloudWatch agent configuration file wizard on each EC2 instance and select the advanced detail level. This will capture the operating system log files.
Answer: A
The short version
A — Add a second CloudWatch agent configuration file and restart the agent with --append-config via SSM Run Command.
Appending merges the new DHCP log stanza with the existing baseline config across all instances without any manual logins.
Key concepts
CloudWatch agent configuration: a JSON file defining which log files (and metrics) to collect; one file can include several logs sections.
--append-config: the agent startup option that merges an additional configuration file with the common configuration already in place instead of replacing it.
AWS Systems Manager Run Command: a one-shot, targeted command execution across many managed instances, with output captured centrally — no SSH and no per-host work.
Windows event/channel logs and file logs: on Windows the agent collects both traditional file logs and specific channels; DHCP information is exposed as a file/log source that must be configured explicitly.
Configuration wizard: an interactive tool that generates a configuration file; running it again on an instance replaces rather than merges the existing configuration.
Why A is correct
The baseline configuration is already deployed and must keep working, so the new DHCP log source has to be added without destroying what exists. A separate configuration file plus --append-config is exactly the mechanism for that: the agent loads both, so the baseline log files and the DHCP logs are collected together. Delivering the file and restarting the agent with the flag through SSM Run Command means the change is applied to all 100 instances as a single auditable, centrally managed operation, which is the definition of operationally efficient. Nothing has to be edited by hand and the wizard never overwrites the running configuration.
Why the others are wrong
B. Logging in to each instance and scripting a push of the log files means 100 separate interventions, credentials in use, and configuration that can drift between hosts. That is precisely the manual operational overhead this answer avoids.
C. Re-running the configuration wizard on each instance regenerates and replaces the agent configuration, which risks losing the carefully deployed baseline settings, and it requires an interactive session on 100 hosts. It also cannot target only the 50 instances that actually have DHCP logs.
D. The "advanced" detail level only selects a broader predefined set of standard operating-system log files. It does not add DHCP-specific sources, and running the wizard again still means 100 interactive sessions that overwrite the deployed configuration.
SOA-C03 exam tip
Agent config rule of thumb: new log source, existing config? Write an additional file and start the agent with --append-config, applied through SSM Run Command. Re-running the wizard replaces your baseline config — never do it in production.
4A company is storing backups in an Amazon S3 bucket. The backups must not be deleted for at least 3 months after the backups are created.
What should a CloudOps engineer do to meet this requirement?
Configure an IAM policy that denies the s3:DeleteObject action for all users. Three months after an object is written, remove the policy.
Enable S3 Object Lock on a new S3 bucket in compliance mode. Place all backups in the new S3 bucket with a retention period of 3 months.
Enable S3 Versioning on the existing S3 bucket. Configure S3 Lifecycle rules to protect the backups.
Enable S3 Object Lock on a new S3 bucket in governance mode. Place all backups in the new S3 bucket with a retention period of 3 months.
Answer: B
The short version
B — Enable S3 Object Lock in compliance mode on a new bucket with a 3-month retention period.
Compliance mode is the only mode in which nobody — not even the root user — can shorten or remove the retention before it expires.
Key concepts
S3 Object Lock: a retention setting at the object/version level that is enforced by S3 itself, rather than by a policy that a privileged user could edit away.
Compliance mode: retention cannot be shortened or deleted by any user, including the account root, for the life of the retention period; only after it expires can the object be deleted.
Governance mode: users with the s3:BypassGovernanceRetention permission can delete or shorten a protected version, so it is a soft guarantee that depends on permissions.
Object Lock requires versioning: Object Lock can only be enabled at bucket creation, on a bucket that has S3 Versioning enabled, which is why the answer specifies a new bucket.
Lifecycle expiration / versioning: a lifecycle rule can delete objects on a schedule, and versioning alone keeps older copies — neither guarantees that a backup survives three months.
Why B is correct
The requirement is a guarantee that backups are not deleted for at least three months. Object Lock in compliance mode is enforced at the service layer and cannot be circumvented: no IAM policy, bucket policy or root action can shorten or bypass the retention period while it is active, which is exactly the strength the wording "must not be deleted" demands. Because Object Lock cannot be added to an existing bucket, the backups must move to a newly created bucket that has versioning and Object Lock enabled, with a default or per-object retention period of three months. Governance mode, IAM denies and lifecycle rules all leave a path — a permitted user, a privileged identity, or the lifecycle rule itself — that removes the data early.
Why the others are wrong
A. A deny on s3:DeleteObject is only as strong as the IAM administration around it: anyone who can edit that policy can delete it, and root or the account owner can always remove a managed or attached policy. It also does nothing about object-level retention or deletion by other means, so it is not a three-month guarantee.
C. Versioning keeps older versions available and lifecycle rules can manage them, but a lifecycle expiration rule that reaches the backups will delete them, and nothing prevents an explicit delete of the current version. Versioning plus lifecycle is a data-management setup, not a retention guarantee.
D. Governance mode is the trap: a user holding s3:BypassGovernanceRetention can remove a protected object before the three months are up. If the requirement is a hard "must not be deleted", only compliance mode satisfies it.
SOA-C03 exam tip
"Backups must not be deleted" = Object Lock compliance mode, enabled at bucket creation with versioning. Governance mode can be bypassed by a privileged user, so it is not a guarantee. Lifecycle rules delete; they never protect.
5A company's CloudOps engineer is troubleshooting communication between the components of an application. The company configured VPC flow logs to be published to Amazon CloudWatch Logs However, there are no logs in CloudWatch Logs.
What could be blocking the VPC flow logs from being published to CloudWatch Logs?
The IAM policy that is attached to the IAM role for the flow log is missing the logs:CreateLogGroup permission.
The IAM policy that is attached to the IAM role for the flow log is missing the logs:CreateExportTask permission.
The VPC is configured for IPv6 addresses.
The VPC is peered with another VPC in the AWS account
Answer: A
The short version
A — The flow log IAM role is missing logs:CreateLogGroup, so the group is never created and the flow log has nowhere to publish.
The logs:CreateLogGroup action must be present on the publishing role in addition to the stream and put permissions.
Key concepts
VPC flow logs: records accepted/accepted-rejected traffic metadata for ENIs, published to CloudWatch Logs, Amazon S3, or Kinesis Data Firehose.
IAM role for flow logs: the role assumed to deliver the records; its policy must allow log-group creation plus logs:CreateLogStream, logs:PutLogEvents and the describe actions.
logs:CreateLogGroup: creates the CloudWatch Logs log group itself; without it the first publish attempt fails and nothing appears.
logs:CreateExportTask: an export job that copies existing CloudWatch Logs log groups to Amazon S3 — unrelated to receiving real-time flow log records.
Flow log format and destination selection: the flow log configuration chooses the destination type and format; the destination type must support CloudWatch Logs for this to apply.
Why A is correct
When flow logs are published to CloudWatch Logs, AWS uses the associated role to create and write to the destination log group. The canonical publishing policy contains logs:CreateLogGroup, logs:CreateLogStream, logs:PutLogEvents, logs:DescribeLogGroups and logs:DescribeLogStreams. If logs:CreateLogGroup is absent, the role cannot create the log group, the delivery attempt is denied, and because flow log delivery failures are silent from the VPC's perspective, the symptom is exactly "the flow log is configured but CloudWatch Logs is empty." Every other permission in the policy may be present and correct while this single missing action still produces nothing.
Why the others are wrong
B.logs:CreateExportTask belongs to exporting existing logs to S3. Flow log delivery to CloudWatch Logs never calls it, so its presence or absence has no effect on whether records arrive.
C. IPv6 addressing does not prevent flow log publication. Flow logs capture IPv4 and IPv6 traffic metadata, and an IPv6 VPC can publish to CloudWatch Logs exactly like an IPv4 one.
D. VPC peering does not block flow logs to CloudWatch Logs. Flow logs are published by the flow log service in the Region, and peering traffic is simply recorded as part of the log content, not blocked from being published.
SOA-C03 exam tip
"Configured but nothing in CloudWatch Logs" is almost always the publishing IAM role. Learn the five actions — CreateLogGroup, CreateLogStream, PutLogEvents, DescribeLogGroups, DescribeLogStreams — and remember CreateLogGroup is the one most often missing.
6A company is migrating a legacy application to AWS. The company manually installs and configures the legacy application on Amazon EC2 instances across multiple Availability Zones. The company sets up an Application Load Balancer (ALB) for the application. The company sets the target group routing algorithm to weighted random. The application requires session affinity.
After the company deploys the application, users report random application errors that were not present in the legacy version of the application. The target group health checks do not show any failures. The company must resolve the application errors.
Which solution will meet this requirement?
Set the routing algorithm of the target group to least outstanding requests.
Turn on anomaly mitigation for the target group.
Turn off the cross-zone load balancing attribute of the target group.
Increase the deregistration delay attribute of the target group.
Answer: A
The short version
A — Set the target group routing algorithm to least outstanding requests.
With session affinity required, weighted random scatters a user's requests across targets that do not share state; least outstanding requests naturally sticks to the target already serving that session.
Key concepts
Routing algorithm: how an Application Load Balancer chooses a target — round robin, weighted random, or least outstanding requests.
Weighted random: distributes requests randomly across the target group, ignoring each target's current load, which is a poor fit when each request must reach the same stateful backend.
Least outstanding requests: routes to the target with the fewest active requests, giving preference to targets with spare capacity and, in a sequential request pattern, effectively sticking a session to one target.
Deregistration delay: how long the load balancer waits for in-flight requests to complete on a target that is being deregistered or is draining.
Why A is correct
The errors appeared only after moving to the ALB, and the health checks are clean, which means the targets themselves are healthy and the fault is in how requests are distributed. The application is stateful: it needs session affinity, so a user whose successive requests land on different instances sees missing session state and throws errors. Weighted random makes that outcome frequent because each request is thrown at an arbitrary target regardless of the previous one. Least outstanding requests selects the target with the fewest active requests, so once a user's request is in flight on a target, subsequent requests for that session continue to favour it while it is busy, giving the sticky behaviour a session-oriented application needs while still balancing load across instances with different capacity.
Why the others are wrong
B. Anomaly mitigation removes targets that show a connection or request anomaly, such as a burst of 5xx responses. The scenario explicitly states health checks show no failures, so there is no anomaly to detect, and removing instances would only reduce capacity further.
C. Cross-zone load balancing controls whether the load balancer sends traffic to targets in other Availability Zones. Turning it off constrains the load balancer to targets in the Zones where targets exist, which reduces availability and capacity without addressing request-to-target distribution or session state.
D. The deregistration delay governs how long existing connections are allowed to drain on a target that is being removed or replaced. The problem here is steady-state routing to healthy targets, so lengthening the drain window changes nothing about which target receives each request.
SOA-C03 exam tip
Stateful app + random errors + healthy targets = wrong routing algorithm, so switch weighted random to least outstanding requests. Anomaly mitigation needs broken targets; deregistration delay only affects draining.
7A company is using an Amazon Aurora MySQL DB cluster that has point-in-time recovery, backtracking, and automatic backup enabled. A CloudOps engineer needs to be able to roll back the DB cluster to a specific recovery point within the previous 72 hours. Restores must be completed in the same production DB cluster.
Which solution will meet these requirements?
Create an Aurora Replica. Promote the replica to replace the primary DB instance.
Create an AWS Lambda function to restore an automatic backup to the existing DB cluster.
Use backtracking to rewind the existing DB cluster to the desired recovery point.
Use point-in-time recovery to restore the existing DB duster to the desired recovery point.
Answer: C
The short version
C — Use backtracking to rewind the existing DB cluster to the chosen recovery point.
Backtracking is an in-place, continuous rewind of the same cluster, whereas point-in-time recovery always produces a new cluster.
Key concepts
Aurora backtracking: continuously records data changes and can rewind the existing cluster to any point within the backtrack retention window, restoring the cluster itself rather than creating a copy.
Backtrack retention period: how far back the cluster can be rewound (24 hours by default, adjustable up to 35 days) — it must be raised to cover 72 hours.
Point-in-time recovery (PITR): restores to a chosen second, but it creates a new DB cluster from the backups and you switch over to it.
Aurora Replica and promotion: a replica can be promoted to become the primary writer of the cluster, which is a failover/rollover operation, not an arbitrary rewind.
Automatic backups: retained backups used for PITR and cluster restores; they cannot be restored over an existing cluster in place.
Why C is correct
Two constraints decide the answer: the recovery point must be within the last 72 hours, and the restore must happen in the same production cluster. Backtracking is the only Aurora feature that satisfies both — it moves the existing cluster's data pages backwards to the chosen timestamp without creating a second cluster or requiring a failover. Because the default backtrack window is 24 hours, the engineer must first raise the cluster's backtrack retention period to at least 72 hours, but the rewind itself is then an in-console operation. Point-in-time recovery fails the second requirement because it produces a new cluster that the application would then have to be pointed at, which is a cutover, not a rollback of the same cluster.
Why the others are wrong
A. Creating a replica and promoting it replaces the primary writer, so it is a promotion/failover path rather than a rewind to a specific point in the last 72 hours. Promotion also carries a data-loss window and changes the cluster's topology, which is not the same operation the scenario describes.
B. AWS Lambda cannot restore a backup into an existing Aurora cluster. A backup restore always produces a new cluster, and a Lambda function is an orchestrator that would still be calling an API that creates a new cluster — it does not make the restore in-place.
D. Point-in-time recovery is the correct choice for restoring to a specific second, but it creates a new cluster from the continuous backups. The requirement that restores be completed in the same production cluster rules it out.
SOA-C03 exam tip
Aurora: backtracking rewinds the same cluster (window = backtrack retention, default 24h); PITR creates a new cluster. Read the words "same cluster" — that decides it.
8A CloudOps engineer is troubleshooting an AWS CloudFormation stack creation that failed. Before the CloudOps engineer can identify the problem, the stack and its resources are deleted. For future deployments, the CloudOps engineer must preserve any resources that CloudFormation successfully created.
What should the CloudOps engineer do to meet this requirement?
Set the value of the DisableRollback parameter to False during stack creation.
Set the value of the OnFailure parameter to DO_NOTHING during stack creation.
Specify a rollback configuration that has a rollback trigger of DO_NOTHING during stack creation
Set the value of the OnFailure parameter to ROLLBACK during stack creation.
Answer: B
The short version
B — Set OnFailure to DO_NOTHING during stack creation.
That value tells CloudFormation to leave the stack in the CREATE_FAILED state with the resources that already succeeded still in place.
Key concepts
OnFailure: a stack-creation option with values ROLLBACK (default) and DO_NOTHING, controlling what happens to already-created resources when the create fails.
ROLLBACK: CloudFormation deletes the resources it created so the failed stack leaves no partial infrastructure behind.
DO_NOTHING: CloudFormation preserves the successfully created resources, and the stack ends in CREATE_FAILED for later investigation and manual cleanup.
DisableRollback: a create-time flag; setting it to False leaves rollback enabled, which is the opposite of what is needed.
Rollback triggers: an alarm-based mechanism evaluated during updates that can trigger rollback; DO_NOTHING is not a valid trigger value in that configuration.
Why B is correct
The engineer's real problem is that the stack and its resources were deleted before the cause could be diagnosed. OnFailure: DO_NOTHING is the documented way to get the opposite behaviour on a create: CloudFormation stops at the failing resource, marks the stack CREATE_FAILED, and deliberately does not delete the resources that were already provisioned, so the engineer can inspect what was built. Nothing needs to be recreated to investigate, and the retained resources can be adopted into a subsequent deployment or cleaned up deliberately.
Why the others are wrong
A.DisableRollback set to False leaves rollback enabled, which is the default behaviour: on failure CloudFormation deletes the partially created resources — exactly the outcome the engineer is trying to avoid.
C. Rollback triggers are for reacting to CloudWatch alarms during an update, and DO_NOTHING is not a valid rollback trigger value, so this configuration is not something CloudFormation will accept as a way to preserve resources.
D.ROLLBACK is the opposite of the requirement: it removes the resources that were successfully created, which is precisely how the previous deployment ended up with a deleted stack.
SOA-C03 exam tip
Debug a failed create → OnFailure: DO_NOTHING keeps the partial resources and the CREATE_FAILED stack. ROLLBACK deletes them. DisableRollback: False keeps rollback on, so it is the wrong flag direction.
9A company plans to run a public web application on Amazon EC2 instances behind an Elastic Load Balancing (ELB) load balancer. The company’s security team wants to protect the website by using AWS Certificate Manager (ACM) certificates. The load balancer must automatically redirect any HTTP requests to HTTPS.
Which solution will meet these requirements?
Create an Application Load Balancer that has one HTTPS listener on port 80. Attach an SSL/TLS certificate to listener port 80. Create a rule to redirect requests from HTTP to HTTPS.
Create an Application Load Balancer that has one HTTP listener on port 80 and one HTTPS protocol listener on port 443. Attach an SSL/TLS certificate to listener port 443. Create a rule to redirect requests from port 80 to port 443.
Create an Application Load Balancer that has two TCP listeners on port 80 and port 443. Attach an SSL/TLS certificate to listener port 443. Create a rule to redirect requests from port 80 to port 443.
Create a Network Load Balancer that has two TCP listeners on port 80 and port 443. Attach an SSL/TLS certificate to listener port 443. Create a rule to redirect requests from port 80 to port 443.
Answer: B
The short version
B — An ALB with an HTTP listener on 80 and an HTTPS listener on 443 carrying the ACM certificate, plus a rule redirecting 80 to 443.
A redirect needs two listeners: the plaintext one that receives the HTTP request and the TLS one it sends traffic to.
Key concepts
Listener: the port plus protocol combination an Application Load Balancer serves — HTTP, HTTPS or gRPC; ALB does not speak raw TCP.
HTTPS listener: requires an SSL/TLS certificate, defaulting to the first certificate attached to the load balancer, which is supplied by AWS Certificate Manager.
Listener rule: matches requests on a listener by condition and forwards them to a target group or performs a fixed action; a redirect action sends the client to another URL or port.
Application Load Balancer: layer 7, HTTP/HTTPS aware, and therefore the only load balancer type that can inspect HTTP and issue protocol redirects.
Network Load Balancer: layer 4, operates at the TCP level, and cannot evaluate HTTP paths or run redirect listener rules.
Why B is correct
Automatic HTTP-to-HTTPS redirection on a load balancer is implemented with two listeners and one rule. The client connects to port 80 and hits the HTTP listener; a listener rule matches and performs a redirect action that returns a 3xx response pointing the client at the https:// equivalent of the request. The HTTPS listener on port 443 then terminates TLS with the ACM certificate and forwards the request to the targets. Because the certificate is attached to the listener that actually serves TLS, the handshake succeeds and browsers show no certificate warning.
Why the others are wrong
A. Declaring an HTTPS listener on port 80 is a configuration mistake: TLS is not served on port 80, so ordinary browser HTTP requests would fail to negotiate, and there is no plaintext listener to receive and redirect the request in the first place. The certificate is attached to the wrong listener.
C. ALB listeners cannot use the TCP protocol — layer 4 TCP listeners are a Network Load Balancer feature. An ALB with TCP listeners on 80 and 443 is not valid, and an HTTP redirect rule cannot be evaluated without terminating HTTP.
D. A Network Load Balancer passes TCP through and has no visibility into HTTP requests, so it cannot match a request path or issue an HTTP redirect. It also lacks the ability to use listener rules for redirection, which is an ALB feature.
SOA-C03 exam tip
ACM + redirect to HTTPS = ALB with HTTP:80 listener + redirect rule → HTTPS:443 listener with the cert. Layer 7 redirects are ALB-only; NLBs and TCP listeners never see the URL.
10A company uses AWS Organizations to manage a set of AWS accounts. The company has set up organizational units (OUs) in the organization. An application OU supports various applications.
A CloudOps engineer must prevent users from launching Amazon EC2 instances that do not have a CostCenter-Project tag into any account in the application OU. The restriction must apply only to accounts in the application OU.
Which solution will meet these requirements?
Create an IAM group that has a policy that allows the ec2:RunInstances action when the CostCenter-Project tag is present. Place all IAM users who need access to the application accounts in the IAM group.
Create a service control policy (SCP) that denies the oc2:RunInstances action when the CostCenter-Project tag is missing. Attach the SCP to the application OU.
Create an IAM role that has a policy that allows the oc2:RunInstances action when the CostCenter-Project tag is present. Attach the IAM role to the IAM users that are in the application OU accounts.
Create a service control policy (SCP) that denies the ec2:RunInstances action when the CostCenter-Project tag is missing. Attach the SCP to the root OU.
Answer: B
The short version
B — Attach an SCP to the application OU that denies ec2:RunInstances when the CostCenter-Project request tag is missing.
SCPs are the only policy type that applies uniformly to every account and principal inside an OU, and a deny cannot be granted away.
Key concepts
Service control policy (SCP): an IAM policy type attached to an OU, account or root that sets the maximum permissions available to everything beneath it.
Maximum permissions model: the effective permissions of a principal are the intersection of its identity policies and every SCP that applies to it; an explicit deny in an SCP cannot be overridden by an allow anywhere.
Tag-based conditions in SCPs: using aws:RequestTag (and aws:TagKeys) in a condition so the deny applies only when the requested resource is missing a specific tag.
AWS Organizations OU: a container that groups accounts so that a single SCP can be applied to all of them, and inherited by child OUs and accounts.
IAM identity policy: an allow attached to a user or role; it can grant access but cannot reliably prevent an action that another permission source also allows.
Why B is correct
The constraint is that every user in every account of the application OU must be prevented from launching an untagged EC2 instance, and only that OU is affected. An SCP attached to the application OU sits above all of those accounts, so it applies to every principal in them regardless of how they authenticate, including federated users, roles and root. Writing it as Deny on ec2:RunInstances conditioned on the CostCenter-Project request tag being absent means the action is evaluated at launch time, where request tags are known: an instance created without the tag is rejected by the API call. Because an SCP deny is a ceiling, a user cannot grant themselves an exception with an IAM allow. Attaching it to the OU rather than the root also keeps the restriction scoped exactly as required.
Why the others are wrong
A. An IAM group policy grants permissions to the users placed in the group; it does not deny anything by default, so a user who happens to have another path to ec2:RunInstances is unaffected. It also has to be maintained in each account where the group is replicated, and it does not apply to roles or federated identities at all.
C. An identity policy on a role expresses an allow, not a guarantee — nothing stops another principal in the account, or a different role, from launching untagged instances. Attaching policies to IAM users in member accounts also does not cover federated or assumed-role access, so it cannot be the OU-wide control the question requires.
D. Attaching the SCP to the root applies the restriction to every account in the organization, not only the application OU. The question explicitly requires the restriction to apply only to accounts in that OU, so the scope is wrong.
SOA-C03 exam tip
OU-wide, non-bypassable restriction → SCP, attached to the OU (root would be organisation-wide). And guardrails against untagged resources: Deny + aws:RequestTag on the create action.