Question 1

How does AWS billing work?

Accepted Answer

Charges accrue per service by usage — compute per second or hour, storage per GB-month, requests per million, and data transfer out per GB. Consolidated billing under an Organizations management account aggregates member accounts and shares volume discount tiers and commitment benefits across them. Two facts worth naming: data transfer in is generally free while out is charged, and cross-AZ traffic within a region is charged in both directions. The second is an invisible cost in multi-AZ architectures that people rarely account for.

Question 2

What is the difference between Savings Plans and Reserved Instances?

Accepted Answer

Reserved Instances commit to a specific instance configuration in a region or AZ. Savings Plans commit to an hourly spend amount: Compute Savings Plans apply across instance family, size, region, tenancy, and to Fargate and Lambda; EC2 Instance Savings Plans are cheaper but locked to a family and region. Compute Savings Plans are the current default recommendation because flexibility protects you against stranded commitment when the architecture changes. Naming that they cover Fargate and Lambda is the detail that makes them attractive to a modernising estate.

Question 3

How do you decide how much to commit to?

Accepted Answer

Analyse the Cost and Usage Report for the stable baseline over several months — the level below which usage never falls — and commit only to that, leaving the variable portion on demand. Model the break-even, check the architecture is not about to change, and prefer one-year terms when uncertainty is high. The failure to avoid is committing to average or peak usage rather than the floor, which leaves you paying for capacity you do not use during troughs. Stating the "commit to the floor" rule explicitly is the concise correct answer.

Question 4

What are Spot Instances and when are they appropriate?

Accepted Answer

Spot uses spare EC2 capacity at up to 90% off, reclaimed with two minutes of notice. They suit fault-tolerant, restartable, stateless workloads — batch processing, CI runners, rendering, data pipelines with checkpointing, and non-critical portions of a web tier behind a load balancer with on-demand baseline. Diversification across instance types and AZs with the capacity-optimized allocation strategy is what actually reduces interruption rates, and naming it shows real usage rather than theory. The design requirement is handling the interruption notice.

Question 5

What are the biggest sources of waste in a typical AWS estate?

Accepted Answer

Over-provisioned EC2 instances and containers; idle non-production environments running overnight and at weekends; unattached EBS volumes and orphaned snapshots from deregistered AMIs; idle Elastic IPs and load balancers; over-retained CloudWatch Logs; S3 noncurrent versions and incomplete multipart uploads with no lifecycle rules; and NAT gateway data processing that a VPC endpoint would eliminate. The ranking matters: right-sizing compute is usually the largest line, but orphaned volumes and snapshots are the easiest immediate wins and reveal a lifecycle problem. The NAT gateway point is the one most candidates miss entirely.

Question 6

What is AWS Compute Optimizer?

Accepted Answer

Compute Optimizer analyses CloudWatch metrics over a lookback period and recommends optimal instance types, Auto Scaling group configurations, EBS volume types, Lambda memory settings and ECS task sizes, flagging over- and under-provisioned resources with projected savings. The caveat is memory: without the CloudWatch agent, memory utilisation is invisible, so recommendations may under-size a memory-bound workload. Installing the agent before trusting the recommendations is the practical prerequisite.

Question 7

What is the Cost and Usage Report and why does it matter?

Accepted Answer

The CUR is the most detailed billing dataset — a row per resource per usage type per hour, with tags, credits, amortised commitment cost and pricing details — delivered to S3 and typically queried with Athena or loaded into a dashboard. Cost Explorer is aggregated; only the CUR answers resource-level questions. The practical requirement is a labelling standard enforced through IaC, because cost attribution is only as good as the tags. Enabling the CUR before you need it matters too, since it is not retroactive.

Question 8

What is the difference between blended, unblended and amortised cost?

Accepted Answer

Unblended cost is what each account was charged at the time. Blended cost averages commitment benefits across accounts in an organisation. Amortised cost spreads an upfront commitment payment evenly across the term rather than showing it all in one month, which is the most useful view for understanding true run rate. Using amortised cost for run-rate analysis is the practical guidance, because an all-upfront Savings Plan otherwise distorts one month enormously. Knowing which view to use for which question is what makes cost analysis credible.

Question 9

What is AWS Budgets and does it stop spending?

Accepted Answer

Budgets sets thresholds on cost, usage, Savings Plans utilisation or Reserved Instance coverage, and alerts at configured percentages of actual or forecast spend. It does not cap spending by default, though Budget Actions can apply an IAM or SCP policy, or stop instances, when a threshold is breached. Budget Actions are the detail that turns budgets from notification into control, and most candidates do not know they exist. The caution is that stopping resources automatically can cause an outage, so actions are appropriate for sandboxes rather than production.

Question 10

What is Cost Anomaly Detection?

Accepted Answer

Cost Anomaly Detection uses machine learning to model normal spend per service, account or cost category and alerts when actual spend deviates significantly, with a root-cause breakdown. It catches sudden changes far sooner than a monthly budget threshold would. Daily anomaly detection is the control that catches a runaway recursive Lambda or an unbounded pipeline within hours rather than at month end. Treating a cost anomaly as an incident with the same urgency as an outage is the operational framing.

Question 11

How do you reduce EC2 costs?

Accepted Answer

Right-size from Compute Optimizer; stop non-production instances outside working hours with Instance Scheduler; move fault-tolerant workloads to Spot; apply Compute Savings Plans to the steady baseline; migrate suitable workloads to Graviton for better price-performance; switch gp2 volumes to gp3; and delete unattached volumes, old snapshots and idle Elastic IPs. The gp2-to-gp3 change is a concrete, low-risk saving of roughly 20% on EBS. Graviton is typically another 20% for interpreted runtimes with no code change, and naming both gives two specific actions rather than general advice.

Question 12

How do you reduce S3 costs?

Accepted Answer

Lifecycle rules tiering to IA, Glacier and Deep Archive; Intelligent-Tiering where access patterns are unknown; expire noncurrent versions; abort incomplete multipart uploads; use S3 Storage Lens to find the largest and least-accessed buckets; and reduce request volume by avoiding repeated LIST operations and caching with CloudFront. Incomplete multipart uploads and noncurrent versions are the two invisible costs, because neither appears in a normal object listing. Adding those two lifecycle rules to every bucket is standard hygiene that most estates lack.

Question 13

How do you reduce data transfer costs?

Accepted Answer

Co-locate compute with data to avoid cross-AZ and cross-region traffic; use VPC gateway endpoints for S3 and DynamoDB to eliminate NAT data-processing charges; serve content through CloudFront, where origin-to-edge transfer is free and cached responses avoid origin egress entirely; and compress responses. The NAT gateway data-processing charge is the one people miss most: every byte a private instance sends to S3 through NAT is charged twice over, and a free gateway endpoint removes it. That is often a surprisingly large line item.

Question 14

How do you reduce CloudWatch and logging costs?

Accepted Answer

Set retention on every log group rather than leaving it infinite; export archives to S3, which is far cheaper; filter out high-volume low-value logs before ingestion; use the Infrequent Access log class for rarely-queried logs; reduce debug logging in production; and prune unused custom metrics and dashboards. Ingestion rather than storage is where most log cost lies, so filtering before it reaches CloudWatch saves more than shortening retention. Log groups left at never-expire are the most common quietly-growing observability cost.

Question 15

How do you reduce serverless costs?

Accepted Answer

Right-size Lambda memory with Power Tuning, since higher memory can reduce total cost for CPU-bound work; use ARM64 Graviton runtimes; shorten duration by removing blocking waits; batch event source records; avoid provisioned concurrency where cold starts are acceptable; and filter events before invocation with EventBridge Pipes or event source filtering. Filtering before invocation is the underrated lever: paying to invoke a function that immediately discards the event is pure waste. That higher memory can be cheaper is the counter-intuitive result worth explaining.

Question 16

How do you implement cost allocation across teams?

Accepted Answer

Prefer account-per-team-per-environment so attribution is structural, supplemented by an enforced tag standard applied through IaC modules and validated with Config rules or tag policies. Activate cost allocation tags, query the CUR by tag and account, and publish per-team dashboards with a weekly summary. Shared costs — networking, logging, a shared cluster — are the hard part and need an explicit allocation key agreed with the teams. Making attribution structural through accounts rather than dependent on tags is what avoids the perennial untagged-resource problem.

Question 17

What is FinOps and what does a good practice look like?

Accepted Answer

FinOps brings financial accountability to variable cloud spend through collaboration between engineering, finance and business: attribution so every cost has an owner, visibility through dashboards, unit economics such as cost per transaction, a commitment strategy reviewed regularly, and optimisation embedded in engineering work rather than run as an annual project. Unit economics is the mature signal: rising total spend is not a problem if cost per transaction is falling. Framing cost as an efficiency metric rather than an absolute is what distinguishes FinOps from cost-cutting.

Question 18

How would you investigate a bill that increased 30% month over month?

Accepted Answer

Query the CUR grouping by service and usage type for both months and compute the delta, then drill into account, tag and resource. Cross-reference with deployment history and Config to see what changed. Cost Explorer's comparison and anomaly views give a fast first pass before the detailed query. Usage-type-level analysis rather than service-level totals is the method, because service totals rarely explain the change. Naming specific likely causes — a new always-on resource, log volume growth, a recursive invocation, cross-region traffic from a deployment change — shows you have done it.

Question 19

How do you prevent a runaway cost incident?

Accepted Answer

Service quotas as hard limits, since budgets only alert; maximum instance counts on autoscaling and Lambda concurrency limits; SCPs restricting expensive instance types and unused regions; Cost Anomaly Detection with daily alerting; and automatic kill switches in sandbox accounts via Budget Actions. The distinction between a quota, which enforces, and a budget, which notifies, is the core of the answer. A recursive Lambda or an unbounded Glue job can spend enormous sums in hours, long before a monthly budget alert matters.

Question 20

When is serverless cheaper than always-on infrastructure, and when is it not?

Accepted Answer

Serverless wins for spiky, low-average-utilisation workloads because you pay nothing when idle — internal tools, event handlers, low-traffic APIs. It loses for high, steady utilisation, where a committed EC2 instance or a well-packed Fargate service costs less per unit of compute. The crossover is the useful part: as utilisation rises, per-request pricing eventually exceeds a reserved instance running continuously. Saying you would model the break-even rather than assert a preference is what makes it an engineering answer.

Question 21

What is the cost impact of Multi-AZ and multi-region?

Accepted Answer

Multi-AZ RDS roughly doubles instance and storage cost because the standby is a full instance, and cross-AZ data transfer is charged both ways. Multi-region adds a full duplicate environment plus cross-region replication and transfer. Both are justified by availability requirements, not applied by default. The concrete recommendation is Multi-AZ in production only — it is often enabled everywhere by a template and doubles non-production cost for no benefit. Challenging whether the stated availability requirement is real is a legitimate engineering move.

Question 22

What is Graviton and why does it matter for cost?

Accepted Answer

Graviton processors are AWS-designed ARM-based CPUs offering meaningfully better price-performance than comparable x86 instances, available on EC2, RDS, Lambda, Fargate, ElastiCache and OpenSearch. Interpreted runtimes and modern compiled languages generally move with no code change. The migration constraint is native dependencies and container base images that must be built for ARM64, plus any x86-only commercial software. Testing before committing is the practical step, and the saving is typically around 20%.

Question 23

How do you cost-optimise a development environment?

Accepted Answer

Schedule shutdown outside working hours; use much smaller instance types and non-HA databases; share a cluster across teams with namespaces and quotas; use Spot where interruption is acceptable; apply aggressive log and data retention; and set hard budgets and quotas with automatic cleanup of resources beyond a maximum age. Turning non-production off at night and weekends is typically a 70% reduction on those environments for workloads nobody uses at 2am. Non-HA databases in non-production is the second largest and equally straightforward saving.

Question 24

What is a service quota and how does it differ from a budget as a cost control?

Accepted Answer

A service quota is a hard technical limit on consumption per account and usually per region, enforced by the platform, so exceeding it fails the request. A budget is a monitoring construct that only notifies. Quotas prevent spend; budgets report it. The practical recommendation is to set quotas deliberately in sandbox and non-production accounts and in regions you do not use, rather than accepting defaults. Zero quota in unused regions doubles as a security control against cryptomining.

Question 25

How do you build a cost dashboard engineering teams will use?

Accepted Answer

From the CUR in Athena or QuickSight, show cost by team and account over time, top movers week over week, cost per environment, unit economics such as cost per thousand requests, and forecast against budget. Push a weekly summary to each team's channel rather than expecting them to visit a dashboard. The behavioural point is that dashboards nobody looks at change nothing — pushing the summary is what creates ownership. Showing week-over-week movers rather than absolute totals is what makes the data actionable.

Question 26

How would you approach a mandate to cut AWS spend by 25% in one quarter?

Accepted Answer

Measure first with the CUR, then work in order of impact and risk: eliminate obvious waste — idle resources, orphaned volumes and snapshots, over-retained logs; schedule non-production shutdown; right-size from Compute Optimizer; move fault-tolerant workloads to Spot and suitable ones to Graviton; apply Savings Plans to the validated baseline; and only then consider architectural change. Track weekly and publish progress. The sequencing is the answer, because it front-loads savings that carry no reliability risk. Committing to a Savings Plan before right-sizing is the classic mistake, since you lock in a baseline you were about to reduce.

Question 27

What is the difference between cost optimisation and cost avoidance?

Accepted Answer

Optimisation reduces the cost of what you already run — right-sizing, commitments, cheaper storage classes. Avoidance prevents cost being incurred at all — designing so idle resources scale to zero, choosing an architecture with a lower cost floor, or deciding not to build something. Avoidance is more valuable but requires influence at design time. The point that follows is that cost is a design property, so a cost estimate in the design document and a cost review in architecture review are cheaper than optimising a running system. Proposing that process is a senior contribution.

Question 28

How do you estimate the cost of a system before building it?

Accepted Answer

Model the dominant drivers rather than every line item: compute hours or request volume, data stored and its growth, data scanned by queries, and network egress. Use the pricing calculator for unit rates, apply expected discounts, present a range with stated assumptions, and validate against actual cost after the first month. Identifying the two or three dominant drivers is what makes an estimate useful, since most services are noise. Committing to validate the estimate against reality is what turns estimation into a skill rather than a guess.

Question 29

What organisational controls support cost governance?

Accepted Answer

SCPs restricting regions and expensive instance types; tag policies requiring cost allocation tags; Control Tower guardrails; service quotas set deliberately per account; budgets with actions in sandbox accounts; and a project-vending pipeline that creates accounts with budgets, quotas and tags already configured. Restricting regions is the underrated one: it prevents accidental deployment into an expensive or non-compliant region and reduces the surface you must monitor. Framing SCPs as a cost control as well as a security control is a useful cross-domain insight.

Question 30

What is the hidden cost of over-provisioning for peak?

Accepted Answer

Capacity sized for peak sits idle most of the time, so cost per unit of useful work is poor. The alternative is autoscaling with a committed baseline: serve the peak on demand and pay discounted rates for the floor, which is usually a large saving against a fixed fleet sized for the busiest hour. The counter-consideration is that autoscaling has a response time, so genuinely spiky traffic may need pre-warming or capacity reservations. Acknowledging that autoscaling is not free of trade-offs is more credible than presenting it as a pure win.

Question 31

What is the cost model difference between Lambda, Fargate and EC2?

Accepted Answer

Lambda bills per request and per GB-second with no idle cost and scales to zero. Fargate bills per vCPU-second and GB-second for allocated task resources, so it costs while tasks run regardless of traffic. EC2 bills per instance-second whether or not it is used, but is cheapest per unit at high utilisation with a Savings Plan. Utilisation determines the winner, and the crossover is where the interesting analysis lies. Adding that EC2 and self-managed Kubernetes carry an operational cost in engineering time — real but absent from the bill — is the observation that impresses.

Question 32

How do you handle Savings Plans in a multi-account organisation?

Accepted Answer

Purchase from the management account so benefits are shared across member accounts by consolidated billing, unless you deliberately restrict sharing. Review utilisation and coverage reports regularly, and treat commitment strategy as a central function while right-sizing stays with the teams. The split of responsibility is the practical point: teams cannot commit meaningfully and a central team cannot right-size individual workloads. Naming that division is what makes a FinOps operating model workable rather than adversarial.

Question 33

What is a Reserved Instance marketplace and can commitments be changed?

Accepted Answer

Standard Reserved Instances can be sold on the marketplace if no longer needed; Convertible RIs can be exchanged for different families. Savings Plans cannot be cancelled, sold or exchanged, so the commitment is firm for its term — which is why committing only to the validated baseline matters. That Savings Plans cannot be unwound is a fact people discover uncomfortably, and it is the strongest argument for conservative commitment levels and one-year rather than three-year terms in a changing estate.

Question 34

How does region choice affect cost?

Accepted Answer

Prices vary by region for compute, storage and egress, sometimes materially. A cheaper region can reduce cost, but must be weighed against latency to users, data residency requirements, service availability — not every service is in every region — and cross-region transfer if the workload talks to resources elsewhere. The caution is that moving compute to a cheap region while leaving data elsewhere often costs more in transfer than it saves in compute. Naming the whole-system view rather than the unit price is what makes the answer correct.

Question 35

What is AWS Cost Categories?

Accepted Answer

Cost Categories group costs by rules — account, tag, service, charge type — into business dimensions such as team, product or cost centre, applied consistently across Cost Explorer, Budgets and the CUR, including for costs that cannot be tagged. They solve the untaggable-cost problem: support charges, shared networking and certain data transfer cannot carry tags, so a rule-based mapping is the only way to attribute them. Naming that gap is what shows practical allocation experience.

Question 36

What is the cost impact of NAT gateways and how do you reduce it?

Accepted Answer

NAT gateways charge hourly per gateway plus per GB processed, and a resilient design needs one per AZ. Reduce it with VPC gateway endpoints for S3 and DynamoDB, which are free and remove that traffic from NAT entirely; interface endpoints for other AWS services; and centralised egress through a shared services VPC in a multi-account estate. The S3 gateway endpoint is the single highest-value change in most estates, because container image pulls, log shipping and data access all flow through NAT otherwise. It is free, improves security, and reduces cost simultaneously.

Question 37

How do you make cost visible during design rather than after deployment?

Accepted Answer

Include a cost estimate in the design document with the dominant drivers modelled; add a cost review question to architecture review; use Infracost or a similar tool in the pull request to show the projected cost delta of an infrastructure change; and set budgets per environment before the first deployment. Showing the cost delta in the pull request is the mechanism that changes behaviour, because it puts the number in front of the engineer at the moment of the decision. That is far more effective than a monthly report after the money is spent.

Question 38

Design a cost governance model for an organisation with 50 teams on AWS.

Accepted Answer

Account-per-team-per-environment under Organizations so attribution is structural, with a vending pipeline creating accounts pre-configured with budgets, quotas, tags and guardrails. CUR to a central account feeding per-team dashboards and a weekly automated summary. SCPs restricting regions and expensive resource types; quotas as the enforcing control. Cost Anomaly Detection with daily alerting treated as an incident class. A central FinOps function owning commitment strategy and shared-cost allocation, with unit-economics metrics per product, while right-sizing and waste elimination stay with the teams. Cost estimates in design documents and cost delta in infrastructure pull requests. The closing scenario. The senior markers are making attribution structural rather than tag-dependent, distinguishing quotas from budgets as the enforcing control, and splitting responsibility so commitment is central while optimisation is local — because neither party can do the other's job.

AWS Cost Optimization & FinOps Interview Questions and Answers

How to use this set