Solutions Architect interview questions
100 real questions with model answers and explanations for Middle candidates.
See a Solutions Architect resume example →Practice with flashcards
Spaced repetition · Hunter Pass
Questions
A horizontally scalable system can add interchangeable instances and distribute work among them.
- Request handling is stateless or stores shared state in external systems available to every instance.
- Load distribution, partitioning, and asynchronous processing prevent one node from receiving all work.
- Dependencies such as databases and queues must scale too, because application replicas cannot remove a downstream bottleneck.
Why interviewers ask this: The interviewer checks whether you understand horizontal scaling as an end-to-end property rather than simply adding servers.
Effective autoscaling combines a meaningful demand signal, a scaling policy, and capacity that can start safely.
- Metrics such as request rate, queue depth, or CPU must correlate with workload pressure and arrive promptly.
- Thresholds, target tracking, cooldowns, and minimum and maximum capacity control scaling behavior.
- New instances need automated configuration, health checks, and enough startup time before receiving traffic.
Why interviewers ask this: A strong answer covers signals, policy, and instance readiness rather than treating autoscaling as a single switch.
Stateful components must preserve ownership, consistency, and movement of data when capacity changes.
- A stateless replica can usually be replaced because durable state lives elsewhere.
- A stateful node may own partitions, sessions, or local data that must be replicated or reassigned.
- Rebalancing and failover therefore require coordination to avoid lost writes, duplicate ownership, or unavailable data.
Why interviewers ask this: The interviewer evaluates whether you connect state to coordination and data movement during scaling.
Load-balancing algorithms choose targets according to a defined view of fairness or current load.
- Round robin rotates through healthy targets, while weighted round robin sends more traffic to higher-capacity targets.
- Least connections prefers the target with fewer active connections and suits requests with uneven duration.
- Consistent hashing maps a key to a stable target and reduces remapping when the target set changes.
Why interviewers ask this: A strong answer distinguishes the mechanics of several algorithms without claiming one is universally best.
Backpressure slows producers when consumers are saturated, while load shedding rejects work that the system cannot process safely.
- Bounded queues, flow-control signals, and reduced intake keep backlog from growing without limit.
- Load shedding fails selected requests early so critical work retains capacity.
- Both techniques protect latency and stability by making overload explicit instead of allowing cascading failure.
Why interviewers ask this: The interviewer checks whether you know how scalable systems remain controlled when demand exceeds capacity.
Cache-aside lets the application load data into the cache only after a cache miss.
- The application checks the cache first and reads the database when the key is absent.
- It then stores the database result with an appropriate TTL before returning it.
- Writes update the source of truth and invalidate or refresh affected cache entries to limit stale data.
Why interviewers ask this: A strong answer describes both the read path and the consistency work required on writes.
Write-through updates the cache and durable store synchronously, while write-behind persists cached writes later.
- Write-through keeps the backing store current but adds its write latency to the request.
- Write-behind can absorb bursts and reduce latency by batching persistence asynchronously.
- Write-behind needs durable buffering and failure handling because unpersisted cache data can otherwise be lost.
Why interviewers ask this: The interviewer evaluates whether you understand the consistency and durability implications of both write strategies.
Cache invalidation uses expiration, explicit removal, or versioned keys to prevent stale entries from living indefinitely.
- TTL-based expiration is simple but permits stale data until the timer ends.
- Explicit invalidation removes or updates known keys when the source changes.
- Versioned keys make new readers use a new namespace while old entries expire naturally.
Why interviewers ask this: A strong answer names concrete invalidation mechanisms and their freshness behavior.
A cache stampede occurs when many requests recompute the same value after an entry expires.
- Simultaneous misses can overwhelm the database or origin with duplicate work.
- Request coalescing or a distributed lock lets one worker refresh the value while others wait.
- Staggered TTLs, early refresh, and serving briefly stale data can keep expirations from aligning.
Why interviewers ask this: The interviewer checks whether you recognize a common failure mode created by otherwise useful caching.
Origin shielding and tiered caching add an intermediate cache layer between edge locations and the origin.
- Many edge misses are consolidated at a regional or shield cache before reaching the origin.
- The design reduces origin connections and protects it during widespread cache misses.
- Cache keys, TTLs, and invalidations must remain consistent across the cache hierarchy.
Why interviewers ask this: A strong answer explains how an additional cache tier reduces duplicated origin traffic.
A read replica is a database copy that receives changes from a primary and serves read traffic.
- Applications direct suitable queries to replicas to reduce read load on the primary.
- Replication lag means a replica may not immediately contain the latest committed write.
- Writes normally remain on the primary unless the database supports a different multi-writer model.
Why interviewers ask this: The interviewer checks whether you understand both read scaling and the consistency effect of replica lag.
Synchronous replication waits for replica confirmation before acknowledging a write, while asynchronous replication does not.
- Synchronous confirmation reduces the chance of losing acknowledged writes during primary failure.
- Its write latency depends on replica response time and network distance.
- Asynchronous replication lowers write latency but can lose recent changes or expose stale reads after failure.
Why interviewers ask this: A strong answer connects acknowledgement timing to latency, consistency, and recovery behavior.
Sharding horizontally divides a dataset across independent database partitions called shards.
- Each shard stores only the rows assigned by a shard key or routing rule.
- Sharding spreads storage and write load beyond the limits of one database node.
- Cross-shard queries, transactions, rebalancing, and hotspot control become application or platform concerns.
Why interviewers ask this: The interviewer evaluates whether you understand both the scaling purpose and the operational complexity of sharding.
A good partition key distributes load and data evenly while supporting common access patterns.
- High cardinality gives enough distinct values to spread records across partitions.
- Even request frequency prevents one popular value from creating a hot partition.
- Queries that include the key can reach one partition instead of scanning or coordinating across many.
Why interviewers ask this: A strong answer links key selection to distribution, hotspots, and query routing.
Connection pooling reuses a bounded set of database connections across many application requests.
- Reuse avoids the latency and resource cost of opening a new connection for every query.
- Pool limits protect the database from more concurrent connections than it can handle.
- Pool size, request timeout, and instance count must be coordinated because every application replica can create its own pool.
Why interviewers ask this: The interviewer checks whether you understand connection count as a shared database capacity constraint.
The CAP theorem says that during a network partition a distributed system must choose between consistency and availability.
- Consistency means every successful read observes the latest committed value under the model being discussed.
- Availability means every request to a non-failing node receives a non-error response.
- Partition tolerance is unavoidable across a real network, so product behavior during partition determines the practical choice.
Why interviewers ask this: A strong answer states the partition condition and avoids the misleading claim that systems freely choose any two properties at all times.
Eventual consistency allows replicas to temporarily differ but expects them to converge when updates stop.
- A read may return an older value while replication or asynchronous processing catches up.
- Conflict resolution and ordering rules determine the result of concurrent updates.
- Systems often add read-your-writes or monotonic-read guarantees where user experience requires them.
Why interviewers ask this: The interviewer evaluates whether you understand eventual consistency as a defined convergence model, not random inconsistency.
CQRS separates models and paths for commands that change state from queries that read state.
- The write model enforces business rules and records accepted changes.
- One or more read models are shaped for efficient queries and may update asynchronously.
- This separation can scale reads and writes independently but introduces synchronization and consistency work.
Why interviewers ask this: A strong answer explains the model separation and the cost of keeping read views current.
Event sourcing stores an ordered history of domain events as the authoritative record of state changes.
- Current state is rebuilt by replaying events or read from projections derived from them.
- The event log provides an audit history and can feed new projections later.
- Event schema evolution, replay behavior, storage growth, and accidental sensitive data require deliberate management.
Why interviewers ask this: The interviewer checks whether you distinguish event sourcing from merely publishing notifications after database updates.
A work queue distributes each message to one competing consumer, while publish-subscribe gives an event to multiple subscriptions.
- Queue consumers share a backlog so adding consumers increases processing capacity.
- Each pub-sub subscription maintains an independent delivery path for its consumer group.
- Both patterns decouple producers, but they express one-to-one work distribution and one-to-many notification respectively.
Why interviewers ask this: A strong answer clearly distinguishes competing consumption from fan-out delivery.
Locked questions
- 21
What do at-most-once, at-least-once, and exactly-once delivery mean?
- 22
What makes a message consumer idempotent?
idempotency - 23
How do partitions affect ordering and scalability in Kafka?
kafkascalabilitypartitioning - 24
What is a dead-letter queue?
data-structures - 25
What is the saga pattern for distributed transactions?
sagadistributedtransactions - 26
What role does an event bus such as AWS EventBridge play?
- 27
What responsibilities belong in an API gateway?
gatewayapi-gatewayapi - 28
What problem does a service mesh solve?
service-mesh - 29
How do Istio's control plane and data plane differ?
- 30
How do REST, gRPC, and GraphQL differ as API styles?
restgraphqlgrpc - 31
How do token-bucket and leaky-bucket rate limiting differ?
tokensrate-limiting - 32
How do timeouts, retries, and circuit breakers work together?
resilience - 33
How do IAM users, roles, and policies relate to one another?
identity-access - 34
What does least privilege mean in IAM design?
least-privilegeidentity-accessdesign - 35
What are identity federation and single sign-on in cloud access?
- 36
How do encryption at rest and encryption in transit differ?
encryptionrest - 37
What is envelope encryption with a cloud KMS?
encryption - 38
What capabilities should a secrets-management service provide?
secrets - 39
How do network segmentation, security groups, network ACLs, and a WAF provide layered security?
networking - 40
What is threat modeling in solution architecture?
threat-modelingarchitecture - 41
What is a multi-tier architecture?
architecture - 42
How do active-active and active-passive multi-region architectures differ?
cloud-regionsmulti-region - 43
What are RTO and RPO in disaster recovery?
disaster-recovery - 44
What are the main cloud disaster-recovery strategies?
- 45
What are the common migration strategies known as the 7 Rs?
migrationscloud-migration - 46
What is the strangler fig pattern for application migration?
migrationcloud-migrationmigrations - 47
What role does change data capture play in data migration?
migrationscloud-migration - 48
What are the main mechanisms for cloud cost optimization?
cost-optimizationoptimizationfinops - 49
How do the six AWS Well-Architected pillars influence a design review?
designwell-architected - 50
How does the AWS Well-Architected review process work?
well-architectedconcurrency - 51
You have six engineers and must launch a B2B SaaS product for 200 tenants in four months, with one tenant expected to generate half the traffic. How would you design tenancy without overbuilding?
designcloud-modelscloud - 52
Design a checkout flow where payment callbacks can be duplicated, inventory is limited, and customers must receive a response within three seconds even if fulfillment is unavailable.
designcallbacks - 53
Users upload files up to 5 GB, processing can take 15 minutes, and the API must respond within two seconds. How would you design the path on AWS?
designapiconcurrency - 54
A marketplace must send email, SMS, and push notifications for two million daily events, honor preferences, and deliver urgent security alerts within one minute. How would you structure it?
alerting - 55
An IoT platform ingests 100,000 events per second with fivefold bursts, must preserve order per device, and needs near-real-time alerts plus cheap seven-year retention. What would you build?
retentionalerting - 56
A web product serves the US, Europe, and Asia, its pages are 90 percent static, and personalized API responses must stay below 300 ms without violating EU data residency. How would you deliver it globally?
discoveryapi - 57
You must integrate with 50 B2B customers using REST, SOAP, and nightly SFTP files, and one customer's outage must not block the others. How would you design the integration layer?
designrest - 58
An ecommerce platform needs strict order and payment consistency, while its catalog has different attributes across thousands of categories and ten times more reads than writes. Would you choose SQL or NoSQL?
sqlnosqlconsistency - 59
A checkout requires a fraud decision before payment, but tax documents, loyalty points, and emails may finish within five minutes. Which interactions should be synchronous and which asynchronous?
async - 60
A team of eight has six months to launch a marketplace with uncertain demand. Would you begin with a monolith or microservices, and how would you preserve scaling options?
microservicesmonolithscaling - 61
A platform needs a public partner API, low-latency internal communication, and a dashboard combining many backend resources. How would you choose among REST, gRPC, and GraphQL?
restgraphqlgrpc - 62
An order platform produces 10,000 events per second, needs per-order ordering and seven-day replay, but the team already operates AWS queues and has no Kafka experience. Would you introduce Kafka?
kafkadata-structures - 63
Support needs a fast cross-order dashboard with filters, order writes require strict validation, and the dashboard may be 30 seconds stale. Is CQRS justified?
validationcqrs - 64
A pricing API receives 50,000 reads per second, prices change every minute, and showing a price older than two minutes is unacceptable. How would you cache it safely?
pricingapicaching - 65
A Kubernetes API normally runs at 20 percent capacity but receives unpredictable eightfold marketing spikes, and CPU does not correlate with request cost. How would you combine load balancing and autoscaling?
load-balancingcapacityscaling - 66
A checkout system needs 99.95 percent availability, RPO under five minutes, users on two continents, and EU orders must remain in Europe. Would you run active-active across regions?
system-designcloud-regions - 67
A PostgreSQL order database reaches 75 percent CPU at peaks, 70 percent of traffic is reads, and traffic should triple within a year. What scaling sequence would you propose?
databasepostgresscaling - 68
Design storage and delivery for private source videos and public transcoded previews, with files from 100 MB to 20 GB and frequent global playback. What belongs in object storage and the CDN?
designcdncloud-storage - 69
Twenty external partners call APIs owned by five teams, each partner needs separate quotas and credentials, and APIs evolve at different rates. What would you place in an API Gateway?
gatewayapi-gatewayapi - 70
A three-person platform team runs 12 Kubernetes services on EKS and is considering Istio for retries, mTLS, and traffic splitting. Would you adopt a service mesh now?
kubernetesservice-meshdecision-making - 71
Two regulated customers must move from a shared SaaS database to dedicated isolation, while 300 standard tenants remain pooled and all receive the same releases. How would you avoid forking the product?
databasecloudcloud-models - 72
You are reviewing document sharing with public links, team permissions, and virus scanning. How would you threat-model it before implementation?
code-review - 73
A regional outage or ransomware incident must not lose more than 15 minutes of order data, and checkout must recover within 60 minutes. What disaster-recovery design would you propose?
designcloud-regionsincidents - 74
Orders are mastered in AWS, CRM runs in Azure, and analytics runs in GCP; updates must arrive within five minutes and customer deletion must propagate everywhere. How would you design the cross-cloud data flow?
design - 75
Three teams are building ordering, inventory, and fulfillment, but the design shares one database and changes frequently break another team's workflow. How would you apply DDD?
databasedesign - 76
An existing AWS application slows down during traffic peaks, but every team blames a different component. How would you find the architectural bottleneck before proposing changes?
trackingarchitecturecomponents - 77
Your AWS workload meets its SLA but cloud spend has grown 45 percent in six months. How would you reduce cost without creating a performance regression?
performance - 78
A team is building an image-processing API with unpredictable bursts, jobs lasting from two seconds to twenty minutes, and a small operations team. Would you choose serverless or containers?
containersapiconcurrency - 79
How would you phase the migration of a revenue-generating on-premises application to AWS when the business cannot accept a big-bang cutover?
migrationscloud-migration - 80
You need to move a busy PostgreSQL database to Amazon RDS with less than five minutes of write downtime. What approach would you defend?
databasepostgres - 81
A modular monolith is becoming hard to release, and one domain causes most incidents. How would you decide whether and how to extract a service?
monolithmicroservicesincidents - 82
Design the integration between an AWS-hosted order service and an on-premises ERP that permits inbound connections only through a private network and has nightly maintenance windows.
design - 83
A critical shipping vendor API has variable latency, duplicate responses, and occasional two-hour outages. How would you protect your application while preserving order flow?
procurementapilatency - 84
A stakeholder requests active-active deployment across AWS and Azure solely to avoid vendor lock-in. How would you evaluate that request?
decision-makingcommunicationdeployment - 85
Your team needs customer identity verification within four months. How would you decide between building it and buying a managed vendor solution?
procurement - 86
Several teams consume Kafka order events, and you need to add a new fulfillment model without breaking older consumers. How would you evolve the event contract?
kafka - 87
An EventBridge rule can deliver the same payment-confirmed event more than once. How would you prevent duplicate downstream fulfillment?
- 88
A product manager wants inventory shown instantly across regions, but the checkout must never oversell scarce items. What consistency model would you propose?
consistencycloud-regions - 89
A customer profile is mastered in one service but must update CRM, billing, and support systems within ten minutes despite partial outages. How would you design the synchronization?
system-designdesign - 90
A public REST API must change its customer representation while mobile clients may remain old for a year. How would you version and roll out the change?
rest - 91
How would you rate-limit a partner API where each customer has a different contract and expensive GraphQL queries can cost far more than simple REST calls?
restgraphqlqueries - 92
What Datadog observability design would you require before launching a new Kubernetes-based checkout service?
kubernetesobservabilitymonitoring - 93
How would you structure Terraform and ArgoCD for three Kubernetes environments without allowing configuration drift or unsafe production promotion?
kubernetesterraformconfig - 94
You are asked to evaluate managed Kafka against a new serverless event service through a proof of concept. What would make the evaluation credible?
kafkaserverlessdecision-making - 95
Business stakeholders ask for a new customer portal but provide only feature requirements. How would you turn missing nonfunctional requirements into an architecture baseline?
stakeholder-managementarchitecturecommunication - 96
Two teams disagree on using REST or gRPC for a latency-sensitive internal service. How would you run and document the decision?
restgrpclatency - 97
A European customer requires personal data to stay in the EU while global support staff still need to troubleshoot the service. How would you shape the solution?
discovery - 98
Cloud users must trigger actions on equipment inside a factory network with intermittent connectivity, and duplicate or stale commands could be unsafe. How would you design the command path?
design - 99
You inherit a costly, fragile application and have two quarters to improve it without stopping feature delivery. What practical architecture roadmap would you propose?
roadmaparchitectureownership - 100
Regulated customers require your SaaS data plane to run in their own AWS accounts while your team operates one shared control plane. How would you avoid bespoke deployments?
deploymentcloudcloud-models