Cloud Infrastructure Interview Questions with Worked Answers
Cloud infrastructure questions connect compute, networking, storage, identity and failure behaviour. Explain how a user request reaches the application, which dependencies can interrupt it and which responsibilities remain with the customer. Then test the design against a concrete failure, such as losing a zone or a required database connection.
Topic: Cloud Computing: AWS, Azure & GCP in IT Infrastructure · Interview preparation
Quick answer: Cloud infrastructure questions connect compute, networking, storage, identity and failure behaviour. Explain how a user request reaches the application, which dependencies can interrupt it and which responsibilities remain with the customer. Then test the design against a concrete failure, such as losing a zone or a required database connection. These original questions and answer cues cover infrastructure reasoning across platforms, with AWS examples where named. They do not represent a provider exam, an employer's private question list or proof that a sample design meets a production availability target.
Key takeaways
- Explain the request path and every dependency it needs.
- Distinguish a service model from a deployment model.
- Choose storage by access and recovery requirements.
- Place replicas across relevant failure domains.
- Verify remaining capacity and dependencies after a failure.
Table of Contents
- What makes infrastructure a cloud service?
- How do VMs, containers and orchestration differ?
- How would you trace a failed application request?
- Which storage choice fits a given workload?
- Do two instances necessarily survive a zone failure?
- How do you distinguish scaling from high availability?
- Who owns security and repeatable infrastructure changes?
- How should you answer the full failure exercise?
- References
What makes infrastructure a cloud service?
Question: is renting one virtual machine enough to explain cloud computing? Answer cue: a VM is a resource, while NIST SP 800-145 describes a service model with five essential characteristics, including on-demand self-service, resource pooling and measured service. Explain the operational arrangement rather than identifying cloud purely by a remote location.
Follow-up: how do IaaS, PaaS and SaaS differ from public, private and hybrid cloud? The first group describes service responsibility; the second describes deployment. An application can use an infrastructure service in a public cloud without becoming a SaaS product. In an interview, name which layer the team operates and which the provider operates. That makes the later security, maintenance and recovery answers more precise.
How do VMs, containers and orchestration differ?
Question: would moving an application into a container remove the need to operate its dependencies? Answer cue: packaging and execution boundaries change, but the application still needs configuration, identity, storage and a deployment process. A VM normally includes a guest operating system, while containers use operating-system isolation rather than requiring a separate guest kernel for each workload.
The Kubernetes overview describes a platform for managing containerised workloads. It is not a replacement for all application or operational work. Ask what happens when a workload restarts: where does its durable data live, how does it regain credentials and when is it ready to accept traffic? A candidate who explains that sequence demonstrates more understanding than someone who says orchestration simply “handles everything automatically.”
How would you trace a failed application request?
Question: a VM is running, but users cannot reach the application. What does that establish? Answer cue: only a compute state. Trace name resolution, destination address, route, policy, listener and application response. Each is a separate possible failure point. A successful test from the VM itself may take a different path from a user request.
The Amazon VPC guide distinguishes networks, subnets, routes and gateways. In its scope, a subnet resides in one Availability Zone. For a fictional internal reporting service, draw the intended client-to-service path and identify the evidence available at each boundary. Do not open a broad firewall rule merely because the final page fails. The interview answer should identify the narrow next question, not change a real network.
Which storage choice fits a given workload?
Question: why not use an object store as though it were a VM's ordinary block device? Answer cue: object, file and block interfaces provide different access semantics. Choose based on application support, sharing, durability, performance and recovery needs. The AWS storage overview describes distinct services rather than one interchangeable storage pool.
For an invented invoice system, uploaded documents may suit object access while its database requires a supported persistence arrangement. Ask how permissions, deletion, retention and restore are handled for each. A durable storage service cannot prevent an authorised application from writing the wrong business value. Similarly, a snapshot's existence does not prove the application can resume correctly. Keep the infrastructure property separate from the end-to-end recovery requirement.
| Requirement | Question to ask |
|---|---|
| VM disk | Which block interface, zone and lifecycle apply? |
| Shared files | Which clients and locking semantics are required? |
| Uploaded objects | How are object access and retention controlled? |
| Database recovery | Which consistent state can the application restore? |
Do two instances necessarily survive a zone failure?
Question: two application instances run in the same zone. Are they protected from losing that zone? Answer cue: their placement shares that failure domain. The AWS Regions and Availability Zones guide explains independent locations within a Region. Application dependencies must also be reviewed.
This offline Python 3 exercise removes one invented zone from two small placement maps. It tests only the map logic; it does not query cloud resources or measure service availability.
layouts = {
"same-zone": {"node-a": "zone-1", "node-b": "zone-1"},
"split-zone": {"node-a": "zone-1", "node-b": "zone-2"},
}
for label, nodes in layouts.items():
survivors = sorted(node for node, zone in nodes.items() if zone != "zone-1")
print(label + ":", ", ".join(survivors) or "none")Expected output:
same-zone: none split-zone: node-b
The split layout retains one node in the model. It still says nothing about database placement, routing, remaining capacity or the application's ability to serve requests. Those are follow-up questions, not facts supplied by the printed result.
How do you distinguish scaling from high availability?
Question: if an application scales out, does it necessarily meet its availability target? Answer cue: scaling changes capacity; availability also depends on placement, health checks, dependencies and failure behaviour. Adding instances behind an unavailable database does not restore the whole service.
For the reporting system, ask whether the surviving node can handle the required workload and whether a healthy path reaches it. Define the traffic and response requirements before claiming headroom. Horizontal scaling may require session or state handling changes; vertical scaling may face limits or interruption. Discuss those application constraints instead of treating either direction as universally better. A useful answer names a measurement that would validate the chosen approach under an approved test.
Who owns security and repeatable infrastructure changes?
Question: does provider-managed infrastructure make customer permissions somebody else's responsibility? Answer cue: apply the AWS shared-responsibility model to the actual service. Physical facilities and a customer's application access are different layers. Identify the data, identity and configuration choices the team still controls.
Question: what does infrastructure as code add? The Terraform introduction describes versionable resource configuration. Reviewable configuration helps explain intended changes, but an approved file alone does not prove the deployed state matches it. In the fictional system, the reviewer needs the target account, proposed network change, approval and observed result. Protect credentials and sensitive state, and keep the change within the team's authority.
How should you answer the full failure exercise?
Bring the pieces together: “The reporting service has two front ends, a database and document storage. I would map where each runs, the client path and the identities used. For a zone failure, I need evidence that a reachable front end and its required data path remain available with enough capacity.” That is an original answer outline, not a tested production design.
Ask the interviewer to change one assumption: a database exists in only the lost zone, a certificate expires, or the remaining node cannot sustain demand. Explain how that changes your conclusion. Keep the accepted service requirement and unresolved dependency visible. This demonstrates infrastructure reasoning without promising a percentage of uptime from a diagram or confusing a deployment tool with an operational result.
| Claim | Evidence still required |
|---|---|
| A node survives | Healthy application and dependency path |
| Two zones are used | Placement of every critical dependency |
| Capacity can grow | Application behaviour and measured limits |
| Configuration is versioned | Approved and verified deployed result |
Related reading: Cloud architecture design tradeoffs; Cloud operations recovery questions.
References
Frequently Asked Questions
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Wow
0
Sad
0
Angry
0