Skip to content

Top 35 Azure Interview Questions and Answers [2026]

October 6, 2026 23 min read
Azure interview guide by Second Talent
TL;DR: Azure interviews in 2026 test the choice between App Service, Container Apps, Functions and AKS, Cosmos DB partitioning, private networking, messaging and managed identities. Candidates should know that Flex Consumption is now the recommended serverless plan for Functions and that support for the .NET in-process model ends on November 10, 2026.

Since October 1, 2025, Azure has required multifactor authentication for create, update and delete calls made through the Azure CLI, PowerShell, the REST API and infrastructure-as-code tools, not just the portal.

Any pipeline still signing in as a user account broke or is about to.

That one change sums up what Azure interviews now probe: identity first, private networking by default, and a steady stream of renames and retirements that a working engineer has to track.

Key takeaways
  1. 1Azure OpenAI and Azure AI Foundry now sit under Microsoft Foundry, and the Assistants API is retired in favor of the Responses API.
  2. 2From Key Vault API version 2026-02-01, Azure RBAC is the default permission model for new vaults, not access policies.
  3. 3Azure Cache for Redis Basic, Standard and Premium retire on September 30, 2028; the replacement is Azure Managed Redis.
  4. 4Kubenet networking in AKS retires on March 31, 2028; Azure CNI Overlay is the migration path.

Compute and Serverless

1. How do you choose between App Service, Container Apps, Functions and AKS?

By how much of the platform the team wants to run. App Service hosts web apps and APIs from code or a container with the least setup.

Container Apps runs containers on a managed Kubernetes platform with scale to zero, revisions and KEDA-based autoscaling, without exposing the Kubernetes API. Functions runs event-driven code billed per execution.

AKS gives the full Kubernetes API and node control, and the team owns upgrades, node pools and add-ons.

Two by two map of Azure compute: Functions on Flex Consumption for event-driven code with little platform control, Functions Premium or Container Apps jobs for event-driven work with more control, App Service and Container Apps for long-running services with little control, and AKS or VMs for full control.

A good answer names the trigger for moving right on that map: needing CRDs, operators, service mesh control or a specific node configuration is a reason for AKS. Wanting "Kubernetes on the CV" is not.

Microsoft's container options comparison makes the same split.

2. Which Azure Functions hosting plan would you pick for a new app?

Flex Consumption, unless there is a reason not to. Microsoft's hosting options page now calls it the recommended serverless plan and labels the original Consumption plan legacy.

Flex Consumption keeps pay-per-execution billing and scale to zero, and adds virtual network integration, a choice of instance memory, always-ready instances to cut cold starts, and per-function scaling up to 1,000 instances.

The alternatives:

  • Premium: pre-warmed instances and longer run times than Consumption, billed for the instances you keep, and it supports containers.
  • Dedicated (App Service plan): runs on an App Service plan you already pay for.
  • Container Apps: Functions packaged as containers next to other microservices, including GPU workloads.
  • Consumption (legacy): still the answer for Windows-only dependencies. Hosting on Linux in this plan retires on September 30, 2028, and gets no new language versions.

3. What is Durable Functions, and what rule must orchestrator code follow?

Durable Functions is an extension that lets you write stateful workflows as code: an orchestrator function calls activity functions, and the runtime checkpoints progress so a workflow can wait for hours or days and survive restarts.

Common patterns are function chaining, fan-out and fan-in, async HTTP APIs with a status endpoint, monitors that poll on a schedule, and waiting for a human approval event.

The rule is determinism. The runtime rebuilds an orchestrator's state by replaying it from its event history, so the code constraints forbid direct I/O, DateTime.Now, random numbers and new GUIDs inside the orchestrator.

Use the context's current time and put all I/O in activities. Microsoft now recommends the Durable Task Scheduler as the storage backend for new apps.

4. What are App Service deployment slots, and what does a swap do?

A slot is a live copy of the app with its own host name, available on Standard, Premium and Isolated plans (Standard allows five). You deploy to a staging slot, test it, then swap it with production.

App Service warms up every instance of the source slot before it switches routing, so users see no cold start and no dropped requests.

The follow-up is slot settings. App settings and connection strings marked as deployment slot settings stay with the slot during a swap, which is how staging keeps pointing at a staging database.

Rollback is a second swap, since the previous version is still running in the staging slot.

5. What is the difference between availability sets and availability zones?

An availability set spreads VMs across fault domains (separate racks, power and network) and update domains inside one datacenter. It protects against hardware failure and planned maintenance, not against losing the datacenter.

Availability zones are physically separate datacenters within a region, each with independent power, cooling and networking. Spreading VMs or a Virtual Machine Scale Set across zones survives a datacenter outage.

Most new designs use zones, plus zone-redundant options on managed services such as Azure SQL, Storage (ZRS) and App Service.

6. How does Azure Container Apps scale, and what is a revision?

Container Apps scales replicas with KEDA rules: HTTP concurrency, CPU or memory, or event sources such as a Service Bus queue, Event Hubs or Kafka. With a minimum of zero replicas, an idle app costs nothing for compute.

A revision is an immutable snapshot of the app's container and scale configuration, created when either changes.

In multiple-revision mode you can split traffic by percentage between revisions for a blue-green or canary release, then move all traffic to the new one.

For batch work, Container Apps jobs run a container to completion on a schedule, on demand or in response to events.

Azure Kubernetes Service

7. What is AKS Automatic, and how does it differ from AKS Standard?

AKS Automatic is a preconfigured AKS cluster where Microsoft manages the system node pools and uses node autoprovisioning to create nodes from workload requests, so there are no node pools to size.

HPA, KEDA and VPA are enabled, networking is Azure CNI Overlay powered by Cilium, ingress goes through the managed application routing add-on, and egress uses a managed NAT gateway.

AKS Standard leaves those choices to you: node pools, autoscaler, CNI and add-ons. Automatic suits teams that want Kubernetes APIs without running the cluster; Standard suits teams that need control over node images, GPUs or networking.

The AKS Automatic overview lists every default side by side.

8. How should a pod in AKS authenticate to Azure services?

With Microsoft Entra Workload ID. The pod's Kubernetes service account is linked to a managed identity or app registration through a federated identity credential.

AKS issues the pod a service account token, and the Azure Identity library exchanges it for an Entra token, so no secret is stored in the cluster.

A candidate who mentions pod-managed identity is working from old material.

Microsoft deprecated the open-source pod-managed identity project on October 24, 2022, archived it in September 2023, and supported the managed add-on only through September 2025.

Both Workload ID and the migration steps are in the workload identity overview.

9. Which network plugin should a new AKS cluster use?

Azure CNI Overlay for most clusters. Nodes get IP addresses from the virtual network subnet, while pods get addresses from a separate private CIDR that does not consume VNet space, so large clusters do not exhaust the subnet.

Azure CNI with pod subnet or node subnet gives pods routable VNet addresses when something outside the cluster must reach pods directly.

Kubenet, the old default, retires on March 31, 2028, and existing clusters must be upgraded to CNI Overlay before then. Pairing the CNI with Cilium as the dataplane adds eBPF-based network policy and observability.

10. How do you upgrade an AKS cluster without downtime?

Upgrade the control plane first, then node pools one at a time.

AKS adds surge nodes, cordons and drains old ones, and respects Pod Disruption Budgets, so every workload needs more than one replica and a PDB that allows at least one pod to be evicted.

A PDB with maxUnavailable: 0 blocks the drain and stalls the upgrade.

A strong answer separates the two kinds of upgrade: Kubernetes version upgrades and node image upgrades, which patch the OS and can run on their own schedule through auto-upgrade channels and planned maintenance windows.

It also mentions checking deprecated APIs before a version jump. For the Kubernetes side of these questions, see our Kubernetes interview guide.

Data and Storage

11. Explain the Cosmos DB consistency levels.

From strongest to weakest: Strong, Bounded staleness, Session, Consistent prefix and Eventual. Stronger levels give fresher reads at the cost of higher write latency and lower availability across regions; weaker levels give the opposite.

  • Strong: reads always return the latest committed write.
  • Bounded staleness: reads lag writes by at most a configured number of versions or time.
  • Session: within one client session, reads honor read-your-writes, using a session token. The most common choice for user-facing apps.
  • Consistent prefix: reads never see writes out of order, but may be behind.
  • Eventual: no ordering guarantee; replicas converge over time.
The session trap. Session consistency depends on the session token. If one user's requests hit app instances that do not share the token, that user can read their own write as missing.

The consistency documentation covers each level's guarantees.

12. How do you choose a Cosmos DB partition key, and what are RUs?

Pick a key with many distinct values that spreads both storage and requests evenly, and that appears in the filter of your most frequent queries.

Each logical partition holds up to 20 GB, so a key like country or status creates a few hot, capped partitions. userId or tenantId usually works; hierarchical partition keys help when one tenant is far larger than the rest.

Request units are the currency of throughput. A point read of a 1 KB item by ID and partition key costs 1 RU; writes, cross-partition queries and large items cost more.

A query without the partition key fans out to every physical partition, which is the most common reason a Cosmos DB bill surprises a team.

13. Compare the DTU and vCore purchasing models for Azure SQL Database.

DTU bundles compute, memory and I/O into one measure with fixed tiers (Basic, Standard, Premium).

It is simple but opaque. vCore lets you choose compute and storage separately, choose hardware generation, and use Azure Hybrid Benefit for existing SQL Server licenses.

Within vCore, the tiers are General Purpose, Business Critical (local SSD and readable replicas) and Hyperscale (storage that grows to many terabytes with fast restores).

Serverless compute auto-pauses an idle database and bills per second, which suits dev databases and spiky workloads. Elastic pools share one pool of resources across many small databases, the usual pattern for per-tenant databases in SaaS.

14. What are the Blob Storage access tiers?

Hot, cool, cold and archive. Hot has the highest storage cost and the lowest access cost. Cool and cold cost less to store and more to read, with minimum retention of 30 and 90 days before early deletion charges apply.

Archive is offline, with a 180-day minimum.

Reading an archived blob means rehydrating it first. Per the rehydration docs, standard priority can take up to 15 hours for objects under 10 GB, and high priority can finish in under an hour at a higher price.

Lifecycle management policies move blobs between tiers by age or last access, which is how most teams control storage cost.

15. What is replacing Azure Cache for Redis?

Azure Managed Redis. Microsoft's retirement FAQ says the Basic, Standard and Premium tiers of Azure Cache for Redis retire on September 30, 2028, and remaining instances are disabled from October 1, 2028.

The Enterprise tiers have their own migration path.

The interview angle is the caching design, which does not change: cache-aside with a TTL, invalidation on write, and a plan for what happens when the cache is empty or down.

A candidate should also say how they would migrate: stand up the new cache, dual-write or warm it, then move reads.

Networking

16. Compare Load Balancer, Application Gateway, Front Door and Traffic Manager.

  • Azure Load Balancer: layer 4 (TCP and UDP), regional, very low latency. For VMs and internal services.
  • Application Gateway: layer 7, regional. Path-based routing, TLS termination and an optional Web Application Firewall. It can also serve as AKS ingress.
  • Azure Front Door: layer 7, global, at Microsoft's edge. Anycast entry points, caching, WAF and failover between regions.
  • Traffic Manager: DNS-based routing across regions. It returns an endpoint and never sees the traffic itself.

A typical global web app uses Front Door in front of regional App Services or Application Gateways. Front Door (classic) is retiring on March 31, 2027, and no longer accepts new profiles, so new work uses the Standard or Premium tier.

17. What is the difference between a private endpoint and a service endpoint?

A private endpoint puts a network interface with a private IP from your VNet in front of one specific resource, such as a single storage account.

Traffic stays on the Microsoft backbone, the resource's public access can be disabled, and on-premises networks reach it over VPN or ExpressRoute. It needs private DNS zones so the resource's name resolves to the private IP.

A service endpoint only extends the subnet's identity to a whole service; the resource keeps its public IP, and the resource's firewall allows traffic from that subnet. It is free and simple but coarser.

Most security reviews now expect private endpoints for data services.

18. An App Service must call a database with no public access and also be private itself. What do you configure?

Two different features, one for each direction. VNet integration handles outbound traffic: the app's calls to the database leave through a delegated subnet and reach the database's private endpoint.

A private endpoint on the App Service handles inbound traffic: clients inside the network reach the app on a private IP, and public access can be turned off.

VNet integration
  • Outbound: the app calls into the network
  • Needs a delegated subnet
Private endpoint on the app
  • Inbound: clients reach the app on a private IP
  • Lets you turn off public access

Candidates often confuse the two, and it is the most useful networking question for App Service roles.

The follow-up is DNS: both sides need private DNS zones linked to the right VNets, or name resolution returns the public address and the call fails.

19. How do NSGs, Azure Firewall and a hub-and-spoke network fit together?

A network security group is a stateful set of allow and deny rules on IP, port and protocol, applied to a subnet or a network interface.

Azure Firewall is a managed, stateful firewall with FQDN filtering, threat intelligence and central logging.

In a hub-and-spoke design, shared services such as the firewall, VPN gateway and DNS sit in a hub VNet, and each workload gets a peered spoke.

User-defined routes send spoke traffic to the firewall, so egress and spoke-to-spoke traffic are inspected in one place, while NSGs still restrict traffic inside each spoke.

Messaging and Events

20. When do you use Service Bus, Event Grid or Event Hubs?

Microsoft's comparison page draws the line between events and messages. Event Grid routes discrete events, lightweight notifications that something happened, to many subscribers.

Event Hubs ingests high-volume streams of events for analysis. Service Bus delivers messages: high-value commands that one consumer must process, with transactions, ordering and dead-lettering.

Decision flow for Azure messaging: a command that one consumer must process goes to Service Bus, a high-volume stream to analyze goes to Event Hubs, and a notification that something happened goes to Event Grid.

A single system often uses all three: Event Grid announces that a blob was uploaded, Service Bus queues the order that must be processed exactly once in business terms, and Event Hubs collects clickstream telemetry.

21. How do Service Bus sessions, peek-lock and dead-letter queues work?

With peek-lock, a receiver locks a message, processes it, then completes it. If the receiver crashes or the lock expires, the message becomes visible again, so delivery is at least once and handlers must be idempotent.

After the maximum delivery count, the message moves to the dead-letter subqueue for inspection.

Sessions give ordered processing per key: messages with the same session ID go to one receiver at a time, in order, while different sessions are processed in parallel.

It is the answer to "process each customer's events in order without a single-threaded consumer". Topics add publish and subscribe, with filters on each subscription.

22. How do Event Hubs partitions and consumer groups work?

An event hub is split into partitions, each an ordered, append-only log. Events with the same partition key land in the same partition, so ordering holds per key, not across the hub. Throughput scales with the partition count.

Each consumer group is an independent view of the stream with its own read positions. Consumers record a checkpoint, usually in Blob Storage, after processing a batch, and resume from it after a restart.

Events are not deleted when read; they expire after the retention period, which lets a new consumer group replay history. Event Hubs also exposes a Kafka endpoint, so existing Kafka clients can connect without code changes.

Identity and Security

23. What are managed identities, and when do you use user-assigned over system-assigned?

A managed identity is an identity in Microsoft Entra ID that Azure manages for a resource, so code gets tokens without any stored secret.

With DefaultAzureCredential from the Azure Identity library, the same code uses the developer's login locally and the managed identity in Azure.

A system-assigned identity is tied to one resource and deleted with it. A user-assigned identity is a standalone resource you can attach to many resources and create before them, so role assignments exist before the app deploys.

Use user-assigned for scale sets, blue-green deployments and anything recreated often, because a new system identity would need its role assignments granted again.

24. How should an app get secrets from Key Vault, and what changed in 2026?

The app authenticates with its managed identity and is granted a data-plane role such as Key Vault Secrets User, scoped to the vault or a single secret.

App Service and Functions can also reference secrets directly in app settings with Key Vault references, so the code reads an environment variable.

The Key Vault RBAC guide says that from API version 2026-02-01, Azure RBAC is the default permission model for new vaults.

Access policies still exist, but switching an existing vault to RBAC invalidates every access policy at once, which causes outages if the equivalent role assignments were not made first.

Better still is avoiding secrets entirely by using managed identity to reach databases and storage.

25. What is the difference between an app registration and an enterprise application?

An app registration is the application object, the global definition of an app in its home tenant: redirect URIs, credentials, exposed API scopes and required permissions.

An enterprise application is the service principal, the local instance of that app in a specific tenant, where users are assigned, consent is recorded and Conditional Access applies.

A multi-tenant SaaS app has one registration in the vendor's tenant and one service principal in each customer tenant. Role assignments on Azure resources are granted to the service principal, not to the registration.

26. How do Azure RBAC and Azure Policy differ?

RBAC controls who can do what: a role assignment grants a security principal a role at a scope (management group, subscription, resource group or resource), inherited downward.

Azure Policy controls what may exist, whoever creates it: allowed regions, required tags, no public IPs on data services, or diagnostic settings deployed automatically.

A senior answer adds the operational pieces: least-privilege built-in roles instead of Owner, Privileged Identity Management for just-in-time elevation, and policy assignments at the management group level so new subscriptions inherit the guardrails.

DevOps and Operations

27. Bicep, ARM templates or Terraform?

Bicep compiles to ARM JSON, supports every Azure resource type and API version on the day it ships, and keeps no state file because Azure Resource Manager is the state. az deployment group what-if previews changes.

Terraform's azurerm provider reached version 5.0 in July 2026; Terraform suits teams managing several clouds or SaaS providers from one workflow, at the cost of a state file to secure and lock.

Deployment stacks add something both lacked on Azure: a stack tracks the resources a Bicep file created, detaches or deletes those removed from the file according to actionOnUnmanage, and can apply deny settings so nobody changes managed resources outside the pipeline.

See our Terraform interview guide for the state side of that trade-off.

28. How do you instrument an app with Application Insights today?

With the Azure Monitor OpenTelemetry Distro, Microsoft's distribution of the CNCF OpenTelemetry standard, available for .NET, Java, Node.js and Python.

It collects requests, dependencies, exceptions, logs and metrics and sends them to a workspace-based Application Insights resource backed by Log Analytics.

The older Application Insights SDKs are the migration target, not the starting point.

In an interview, expect questions on sampling to control cost, querying with KQL, correlating a request across services through the W3C trace context, and alerting on failure rate rather than individual errors.

29. How would you cut the Azure bill for a running system?

Measure first with Cost Management, grouped by resource group and tag, and set budgets with alerts. Then work down the usual list:

  • Right-size or shut down idle VMs and databases; use serverless or auto-pause for dev environments.
  • Buy reservations or a savings plan for steady compute, and use Spot VMs for interruptible batch work.
  • Move cold data to cheaper blob tiers with lifecycle rules, and set log retention and sampling on Application Insights.
  • Check egress and NAT gateway charges, which grow quietly with chatty cross-region designs.
  • Apply Azure Hybrid Benefit where existing Windows Server or SQL Server licenses allow it.

30. How would you design an Azure app to survive a region outage?

Start from the recovery time and data loss targets, because they decide the cost.

Active-passive uses a paired or chosen secondary region with infrastructure deployed from the same Bicep or Terraform code, data replicated asynchronously (Azure SQL failover groups, geo-redundant storage, Cosmos DB multi-region), and Front Door or Traffic Manager to shift traffic.

Active-active runs both regions at once and needs a data layer that accepts writes in both, such as Cosmos DB with multi-region writes and a conflict policy.

The part most candidates skip: testing the failover. A plan that has never been run, including DNS, secrets and identities in the second region, is a hope, not a design.

What Changed Recently

Oct 1, 2025
Mandatory MFA, phase 2: Azure CLI, PowerShell, REST API and IaC tools
After Mar 31, 2026
Private subnets by default for new virtual networks on new API versions
Nov 10, 2026
Functions in-process model support ends
Mar 31, 2027
Front Door (classic) retires
Mar 31, 2028
AKS kubenet retires
Sep 30, 2028
Linux Consumption plan for Functions and Azure Cache for Redis (Basic, Standard, Premium) retire

31. What happens to .NET Azure Functions on the in-process model?

Per Microsoft's migration guide, support for the in-process model ends on November 10, 2026. Apps must move to the isolated worker model, where functions run in a separate .NET process from the Functions host.

The isolated model decouples the app's .NET version from the host's, supports standard dependency injection and ASP.NET Core integration for HTTP triggers, and allows middleware.

Migration means changing the project's packages from Microsoft.Azure.WebJobs.* to Microsoft.Azure.Functions.Worker.*, adding a Program.cs host builder and updating trigger and binding types, then testing each function, because binding behavior differs in places.

32. How does mandatory MFA affect automation on Azure?

Microsoft's mandatory MFA page describes two phases. Phase 1, from October 2024, requires MFA to create, update or delete resources through the Azure portal and admin centers.

Phase 2, from October 1, 2025, extends it to the Azure CLI, Azure PowerShell, the Azure mobile app, the REST API, the SDKs and infrastructure-as-code tools. Tenants could postpone phase 2 to July 1, 2026 at the latest.

Workload identities are not affected.

The fix for a pipeline that signs in with a user's password is a service principal or managed identity, ideally using workload identity federation from GitHub Actions or Azure DevOps so no client secret exists.

33. Why can a new VM no longer reach the internet by default?

Azure is retiring default outbound access. According to the default outbound access docs, virtual networks created with API versions released after March 31, 2026 get private subnets, with defaultOutboundAccess set to false.

VMs there need an explicit outbound method: a NAT gateway, a load balancer outbound rule, a public IP, or a firewall reached through a user-defined route.

Existing virtual networks are unchanged, and deployments pinned to older API versions keep the old behavior.

A candidate should know the symptom: Windows activation, package installs and updates fail on a new VM until outbound connectivity is configured.

34. What is Microsoft Foundry, and what happened to the Assistants API?

Microsoft Foundry is the current name for the platform previously called Azure AI Studio and then Azure AI Foundry, and Azure AI Services are now Foundry Tools.

An Azure OpenAI resource can be upgraded to a Foundry resource while keeping its endpoint and keys, per the Foundry overview.

The Assistants API is retired. Its replacement is the Responses API, which combines chat completions and the Assistants features, with Foundry Agent Service for hosted agents.

Threads, messages and runs become conversations, items and responses, and new code calls the stable /openai/v1/ routes instead of monthly api-version values. Authentication should use Entra ID and a managed identity rather than API keys.

35. Which other retirements should an Azure developer be tracking?

The ones on the timeline above, with their migration paths:

  • Front Door (classic) on March 31, 2027: migrate to Front Door Standard or Premium; classic already blocks new profiles and managed certificates.
  • AKS kubenet on March 31, 2028: upgrade clusters to Azure CNI Overlay.
  • Linux Consumption plan on September 30, 2028: move function apps to Flex Consumption.
  • Azure Cache for Redis Basic, Standard and Premium on September 30, 2028: move to Azure Managed Redis.

The skill being tested is not memorizing dates. It is knowing where to find them (Azure Updates and Service Health retirement notices) and having a habit of checking before a design review.

Signs of a Strong Answer

  • They default to managed identity and never mention storing a client secret unless asked why not.
  • They separate inbound and outbound private networking for App Service without prompting.
  • They pick a Cosmos DB partition key from the query pattern and can estimate the RU cost of a cross-partition query.
  • They use current names (Microsoft Entra ID, Microsoft Foundry, Flex Consumption, Azure Managed Redis) and know what each replaced.
  • They choose between Service Bus, Event Grid and Event Hubs by message versus event, not by throughput alone.
  • They have run a failover or a slot swap rollback and can say what went wrong the first time.

Hiring Azure Developers

Azure engineers who can design private networking and identity, not just deploy an App Service, are in short supply.

Second Talent matches companies with pre-vetted cloud engineers and DevOps engineers from Asia, screened with questions like these.

Tell us the stack and we send a shortlist within 24 hours. Start hiring, or compare with our AWS and Google Cloud interview guides.

Ready to hire? Hire Cloud Engineers: vetted AI-native talent, with first profiles in 24 hours.

Get developers who ship with AI agents.

Starts at $25/Hour.

Start Hiring

$0 cost until you hire