Question 1

How is billing structured in GCP?

Accepted Answer

A Cloud Billing account is a payment profile that sits outside the resource hierarchy and is linked to one or more projects. Costs accrue per project against the linked billing account. Billing accounts have their own IAM roles, support budgets and alerts, and can export detailed usage and cost data to BigQuery for analysis. The fact to volunteer is that a project with no active billing account has most services disabled, and that moving a project between billing accounts is possible but changes where its costs land. Both matter when reorganising an estate.

Question 2

What is a budget and does it stop spending?

Accepted Answer

A budget sets an amount and triggers alerts at configured percentage thresholds, based on actual or forecast spend. It does not cap spending — exceeding a budget changes nothing by itself. To actually stop spend, route the budget notification to Pub/Sub and have a Cloud Function disable billing on the project. This is one of the most commonly misunderstood facts in GCP and interviewers ask it deliberately. The follow-up caution is that disabling billing terminates resources and can destroy data, so it is appropriate for a sandbox and dangerous for production.

Question 3

What is billing export to BigQuery and why is it essential?

Accepted Answer

Billing export writes detailed usage, cost, credit and pricing data into a BigQuery dataset, with a row per SKU per resource per day, including labels and project. It is essential because the console reports are aggregated — only the export lets you answer questions like "which bucket, which query, which team" and build chargeback. The practical requirement is a labelling standard applied through organisation policy or Terraform, because cost attribution is only as good as the labels. Enabling the export before you need it also matters, since it is not retroactive.

Question 4

What is the difference between sustained use discounts and committed use discounts?

Accepted Answer

Sustained use discounts apply automatically with no commitment, increasing as an eligible VM runs for a larger fraction of the month, up to around 30% on applicable machine families. Committed use discounts require a one or three year commitment to a quantity of resources or an amount of spend, in exchange for roughly 37% and 55% reductions. The critical distinction is that a CUD is a billing commitment, not a capacity reservation — it bills whether or not you use it, and it does not guarantee resources are available. Capacity guarantees require a reservation, which can be combined with a CUD.

Question 5

What is the difference between resource-based and spend-based committed use discounts?

Accepted Answer

A resource-based commitment covers a specific quantity of vCPU and memory in a specific region and machine family, at a deeper discount but with less flexibility. A spend-based commitment covers a dollar amount per hour across eligible services, is more flexible about what you run and where, but discounts less. The choice depends on how stable your architecture is: a mature, unchanging fleet suits resource-based commitments, while a team migrating between machine families or services should prefer spend-based to avoid stranded commitments. Naming stranded commitment as the risk is the mark of experience.

Question 6

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

Accepted Answer

Over-provisioned VMs and GKE nodes running far below their requested capacity; idle VMs left running in non-production; orphaned persistent disks and snapshots from deleted instances; unattached reserved static IPs; over-retained logs and unbounded BigQuery table growth; noncurrent object versions with no lifecycle rule; and always-on development environments that nobody uses at night or weekends. The ranking matters: right-sizing compute is usually the largest line, but orphaned disks and unattached IPs are the easiest immediate wins and often reveal an underlying lifecycle problem. Naming Active Assist recommendations as the systematic way to find them is a strong close.

Question 7

What is Active Assist and what recommendations does it provide?

Accepted Answer

Active Assist is the family of intelligent recommendations across GCP: machine-type right-sizing based on observed utilisation, idle VM and idle persistent disk detection, unattached IP address detection, IAM role recommendations for least privilege, committed use discount recommendations, and BigQuery slot and partitioning suggestions. The caution to apply is the same for all of them: recommendations are based on an observation window, so a workload that spikes quarterly may be recommended for downsizing incorrectly. A review step, and excluding known-periodic workloads, is the responsible way to apply them.

Question 8

How do you reduce Compute Engine costs specifically?

Accepted Answer

Right-size machine types from recommendations; shut down non-production instances outside working hours with a scheduler; move restartable batch to Spot VMs at 60–91% off; apply committed use discounts to the steady baseline while serving peaks on demand; prefer efficient general-purpose families such as E2 where the workload allows; and move boot disks from pd-ssd to pd-balanced where IOPS is not the constraint. The framing that scores best is baseline versus burst: commit to the floor, autoscale the peak, and use Spot for anything fault-tolerant. Adding that you would measure with the billing export before and after turns the list into an engineering process.

Question 9

How do you reduce BigQuery costs?

Accepted Answer

Partition and cluster large tables and enable require_partition_filter; avoid SELECT * and select only needed columns; use dry runs before expensive queries; set custom quotas on daily bytes billed per user and per project; build materialised views and scheduled aggregates so dashboards do not scan raw data; set table and partition expiration; evaluate physical storage billing; and move steady workloads to slot reservations. The single highest-leverage control is a per-user daily bytes-billed quota, because it caps the damage from one runaway query. Most other measures reduce average cost; the quota removes the tail risk, which is the different kind of protection interviewers look for.

Question 10

How do you reduce GKE costs?

Accepted Answer

Right-size pod requests using Vertical Pod Autoscaler recommendations, since unused requested capacity is pure waste; enable cluster autoscaling with sensible minimums and scale node pools to zero where possible; use Spot node pools for fault-tolerant workloads; consolidate small clusters; apply committed use discounts to the baseline; and consider Autopilot so you pay for pod requests rather than node capacity. The key insight is that in Standard mode you pay for nodes, not pods, so bin-packing efficiency is the whole game — and mis-set requests are the usual reason it is poor. GKE cost allocation, which breaks spend down by namespace and label, is the tool that makes it visible.

Question 11

How does network egress pricing work and how do you reduce it?

Accepted Answer

Ingress is free. Traffic within a zone on internal IPs is free; between zones in a region and between regions is charged at increasing rates; egress to the internet is charged by destination and volume, with Premium Tier costing more than Standard. Reduce it by co-locating compute with data, caching at the edge with Cloud CDN, compressing responses, and using Standard Tier for bulk non-user-facing transfer. The cost most teams overlook is internal cross-region traffic — a service calling a database in another region generates continuous egress that never appears as an obvious line item. Naming that internal pattern rather than only internet egress shows you have actually read a bill.

Question 12

What is the difference between a credit, a discount and a promotion in billing data?

Accepted Answer

In the billing export, credits are line items that reduce cost — sustained use discounts, committed use discount credits, free tier allowances, promotional credits and support agreements. Understanding which credits apply is necessary to compute true net cost, because gross usage cost and what you actually pay can differ substantially. The practical consequence is that any cost analysis must sum cost plus credits rather than cost alone, otherwise savings from existing commitments are invisible and you may double-commit. That is a real and common analysis error.

Question 13

How do you implement chargeback or showback across teams?

Accepted Answer

Enforce a labelling standard — team, environment, cost-centre, application — through Terraform modules and organisation policy, or use a project-per-team structure so attribution follows the project. Export billing to BigQuery, build a model that allocates shared costs (networking, logging, shared clusters) by an agreed key, and publish dashboards per team. Shared cost allocation is the hard part and the part interviewers probe: a shared GKE cluster or a shared VPC has real costs that no single team owns. Naming an explicit allocation key — namespace resource usage, or an agreed percentage — is what makes a chargeback model workable rather than contested.

Question 14

What is the GCP free tier and what are its limits?

Accepted Answer

Two parts: a 90-day trial credit for new customers, and the Always Free tier giving perpetual monthly allowances on around 20 products — one e2-micro VM in specific US regions, 5 GB regional Cloud Storage, 2 million Cloud Functions invocations, 1 TB of BigQuery query processing and more. The region restriction is the detail that catches people: an e2-micro in an Indian region is not free. Also, free tier allowances are per billing account, not per project, so several projects share one allowance.

Question 15

How would you find out why last month's bill increased by 30%?

Accepted Answer

Query the billing export in BigQuery, grouping cost by service and SKU for both months and computing the delta, then drill into the largest movers by project, label and resource. Cross-reference with deployment history and Cloud Asset Inventory to see what changed. The console's cost breakdown reports give a fast first pass before the detailed query. The method is what is being tested — SKU-level delta analysis rather than guessing. Naming the specific grouping (service, then SKU, then resource) shows you have actually done it, because service-level totals rarely explain the change on their own.

Question 16

What is the difference between list price, effective price and net cost?

Accepted Answer

List price is the published rate. Effective price is what you pay after automatic discounts such as sustained use, committed use credits, and any negotiated custom pricing. Net cost is effective cost after all credits including promotional and free-tier allowances have been applied. For any large customer there is usually a negotiated agreement, so public calculator figures overstate cost. Knowing that enterprise pricing is negotiated — and that commitments are a lever in that negotiation — is a commercially aware answer.

Question 17

How do you prevent a runaway cost incident?

Accepted Answer

Set quotas — regional CPU, API rate, BigQuery bytes billed per user — because quotas are hard limits while budgets are only alerts. Cap maximum instances on autoscaling services. Set maximum worker counts on data pipelines. Alert on daily anomalous spend rather than monthly totals. Use organisation policy to restrict expensive resource types and regions. Automate a kill switch for sandbox projects. The distinction between a quota (enforced) and a budget (informational) is the core of the answer. A recursive Cloud Function or an unbounded Dataflow job can spend enormous sums in hours, long before a monthly budget alert would matter.

Question 18

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 VM or a well-packed GKE node costs less per unit of compute than per-request pricing. The crossover point is the useful part of the answer: as utilisation rises, per-request pricing eventually exceeds the cost of a reserved instance running continuously. Being able to say you would model that break-even rather than assert a preference is what distinguishes a real answer.

Question 19

What is a Cloud Storage lifecycle policy's role in cost management?

Accepted Answer

It automatically transitions objects to cheaper storage classes as they age and deletes them when retention expires, so storage cost falls without manual intervention. Without lifecycle rules, buckets grow indefinitely and noncurrent versions accumulate invisibly. The trap to name is minimum storage duration: moving data to Archive and deleting it a week later still bills 365 days. Autoclass is the alternative when access patterns are unpredictable, because it waives early-deletion and retrieval fees for objects it manages.

Question 20

How do you cost-optimise a development and testing environment?

Accepted Answer

Schedule shutdown outside working hours with an instance scheduler; use much smaller machine types and non-HA database configurations; share a single cluster across teams with namespaces and quotas; use Spot VMs where interruption is acceptable; apply aggressive log and data retention; and set hard project-level quotas and budgets with automatic cleanup of resources beyond a maximum age. The largest single saving is usually turning non-production off at night and weekends — roughly a 70% reduction on those environments for a workload nobody uses at 2am. Non-HA databases in non-production is the second, since HA roughly doubles the cost.

Question 21

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

Accepted Answer

FinOps is the discipline of bringing financial accountability to variable cloud spend through collaboration between engineering, finance and business. In practice: labelling and attribution so every cost has an owner; visibility through billing export 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: total spend rising is not a problem if cost per transaction is falling. Framing cost as an efficiency metric rather than an absolute number is what separates FinOps from cost-cutting.

Question 22

What are quotas and how do they differ from budgets as a cost control?

Accepted Answer

A quota is a hard technical limit on consumption — regional CPUs, API requests per minute, BigQuery bytes billed per day — 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 rather than accepting defaults, particularly in sandbox and non-production projects and in regions you do not intend to use. Keeping unused regions at zero quota also doubles as a security control against cryptomining.

Question 23

How do you evaluate whether to buy a committed use discount?

Accepted Answer

Analyse the billing export for the stable baseline of usage over the past 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 that the architecture is not about to change in a way that strands the commitment, and prefer shorter terms when uncertainty is high. The failure to avoid is committing to peak or average 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 24

What is the cost impact of choosing a multi-region rather than a regional configuration?

Accepted Answer

Multi-region storage and databases cost more per unit and add cross-region replication traffic, and multi-region Spanner in particular carries a substantial premium. The benefit is higher availability and lower global read latency. The design question is whether the availability target genuinely requires it. The point to make is that multi-region is often chosen reflexively for a single-country application, paying a permanent premium for a capability nobody needs. Asking what the actual availability requirement is, before choosing, is the engineering move.

Question 25

How does labelling support cost management, and what standard would you enforce?

Accepted Answer

Labels attach key-value metadata to resources and appear in the billing export, enabling cost grouping by team, environment, application and cost centre. Enforce a minimum set through Terraform modules so nothing is created without them, and audit compliance with Cloud Asset Inventory queries. The limitation to acknowledge is that not every resource type supports labels and not every cost is attributable — shared networking and logging in particular — so a project-based attribution model is often a necessary complement rather than an alternative.

Question 26

What is the cost difference between Cloud Run, GKE and Compute Engine for the same service?

Accepted Answer

Cloud Run bills per request-second with scale to zero, so a low-traffic service costs almost nothing and a high-traffic one costs more per unit of compute. GKE bills for node capacity whether or not pods use it, so it is cheap only when well packed. Compute Engine bills for the VM continuously, and is cheapest per unit when it runs at high utilisation with a committed use discount. The crossover framing is the answer: utilisation determines the winner. Adding that GKE and Compute Engine carry a real operational cost in engineering time — which does not appear on the bill but is often larger — is the observation that impresses.

Question 27

How would you build a cost dashboard for engineering teams?

Accepted Answer

From the BigQuery billing export, build views that show cost by team label and project over time, top movers week over week, cost per environment, unit economics such as cost per thousand requests, and forecast against budget. Surface it in Looker Studio, refresh daily, and send an automated weekly summary to each team. The behavioural point is that dashboards nobody looks at change nothing — pushing a weekly summary to the team's channel is what creates ownership. Showing week-over-week movers rather than absolute totals is also what makes the data actionable.

Question 28

What is a project-per-environment versus project-per-team model, and how does it affect cost?

Accepted Answer

Project-per-environment separates dev, staging and production, giving clean isolation and easy environment-level cost reporting. Project-per-team gives clean team attribution. Most enterprises combine both — a project per team per environment — which makes cost attribution straightforward without relying on labels. The trade-off is project sprawl and quota management overhead as the estate grows into the hundreds of projects. Naming a project-vending pipeline that creates projects with standard policies, labels and networking is the operational answer to that sprawl.

Question 29

How do you handle the cost of logging and monitoring?

Accepted Answer

Exclude high-volume, low-value logs at ingestion — health checks, successful load-balancer entries; set short retention on the default log bucket and route archives to Cloud Storage; enable Data Access audit logs and VPC Flow Logs selectively rather than everywhere, with sampling; and prune unused custom metrics, which have their own cost. The point to lead with is that ingestion, not retention, is where log cost is incurred — so an exclusion filter saves far more than shortening retention. That is a specific, actionable fact many candidates get backwards.

Question 30

What is Spot VM pricing and what is the risk?

Accepted Answer

Spot VMs use spare capacity at 60–91% below on-demand, with variable pricing and the risk that Compute Engine reclaims the instance at any time with a 30-second notice. The risk is workload interruption, so they are only appropriate for fault-tolerant, restartable, checkpointed work. The design requirement is handling the preemption signal and checkpointing progress, otherwise you lose work rather than money. Also worth naming: Spot capacity in a specific zone can be unavailable entirely, so a multi-zone strategy is needed for reliable throughput.

Question 31

How do you decide between optimising cost and optimising engineering time?

Accepted Answer

Compare the annual saving against the engineering cost to achieve and maintain it. A change saving a small monthly amount but requiring ongoing operational attention is usually a net loss. Prioritise structural savings — right-sizing, commitments, lifecycle rules — that are one-time changes with recurring benefit, over micro-optimisations that add complexity. This is a judgement question and the mature answer explicitly values engineering time. Saying you would not move off a managed service to save a modest sum, because the operational burden costs more, is a stronger answer than maximising savings unconditionally.

Question 32

What is the cost of a Cloud SQL HA configuration, and when is it justified?

Accepted Answer

HA roughly doubles the compute and storage cost because a full standby instance runs continuously in another zone. It is justified for production workloads where a zone failure would be a business incident, and is rarely justified in development or staging environments. The specific recommendation — HA in production only — is one of the easiest concrete savings in a typical estate, because HA is often enabled everywhere by a template. Naming a default that should be environment-specific is a good practical observation.

Question 33

How does the choice of region affect cost?

Accepted Answer

Prices vary by region for compute, storage and network egress, sometimes by a significant margin. Choosing a cheaper region can reduce cost, but must be weighed against latency to users, data residency requirements, and cross-region egress 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 egress than it saves in compute. Naming the whole-system view rather than the unit price is what makes the answer correct.

Question 34

What organisation policy constraints support cost governance?

Accepted Answer

constraints/gcp.resourceLocations to restrict which regions may be used; constraints/compute.vmExternalIpAccess to prevent public IPs, which also limits egress exposure; machine-type restrictions to prevent very large or GPU instances in non-production; and disabling expensive service types in sandbox folders. 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 org policy as a cost control as well as a security control is a useful cross-domain insight.

Question 35

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

Accepted Answer

Measure first with the billing export, then work in order of impact and risk: eliminate obvious waste — idle resources, orphaned disks, unattached IPs, over-retained data; schedule non-production shutdown; right-size from Active Assist recommendations; move fault-tolerant workloads to Spot; apply commitments to the validated baseline; and only then consider architectural change, which is slower and riskier. Track weekly against the target and publish progress. The sequencing — waste, then right-sizing, then commitments, then architecture — is the answer, because it front-loads savings that carry no risk to reliability. Committing to a discount before right-sizing is the classic mistake, since you lock in a baseline you were about to reduce.

Question 36

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 from 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 interviewers want is that cost is a design property. Reviewing cost during architecture review, and putting a cost estimate in a design document, is far cheaper than optimising a running system, and proposing that process is a senior contribution.

Question 37

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

Accepted Answer

Model the dominant cost drivers rather than every SKU: expected compute hours or request volume, data stored and its growth rate, data scanned by queries, and network egress. Use the pricing calculator for unit rates, apply expected discounts, and build a range rather than a single figure. Then validate against actual cost after the first month and correct the model. The discipline of identifying the two or three dominant drivers is what makes an estimate useful — most SKUs are noise. Committing to validate the estimate against reality is what turns estimation into a skill rather than a guess.

Question 38

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

Accepted Answer

Capacity sized for peak is idle most of the time, so utilisation and therefore cost-efficiency are poor. On GCP the alternative is autoscaling with a committed baseline, which serves peak on demand and pays discounted rates for the floor — usually a large saving compared with a fixed fleet sized for the busiest hour. The counter-consideration to acknowledge is that autoscaling has a response time, so genuinely spiky traffic may need pre-warming, minimum instances or a reservation. Saying that autoscaling is not free of trade-offs is more credible than presenting it as a pure win.

Question 39

What billing IAM roles exist and who should have them?

Accepted Answer

Billing Account Administrator manages the billing account and its associations; Billing Account User can link projects to it; Billing Account Viewer can see costs; Billing Account Costs Manager manages budgets and exports. Finance and a small platform group should hold administration; engineering teams generally need viewer access to their own costs. The separation to highlight is that billing roles are independent of project roles, so a project owner cannot necessarily see or change billing. That separation is deliberate and is a legitimate separation-of-duties control worth naming.

Question 40

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

Accepted Answer

Folder structure per business unit and environment with project-per-team-per-environment so attribution is structural rather than label-dependent, supplemented by an enforced label standard from shared Terraform modules. Billing export to BigQuery feeding per-team dashboards and a weekly automated summary. Budgets with alerts at multiple thresholds, and hard quotas plus organisation policy region and machine-type restrictions as the real controls. A central FinOps function owning commitment strategy and shared-cost allocation, with unit-economics metrics per product. Cost review as a standing item in architecture review, and Active Assist recommendations triaged monthly by each team rather than centrally. The closing scenario. The senior markers are making attribution structural rather than relying on labels, distinguishing quotas from budgets as the enforcing control, and pushing optimisation ownership to the teams while keeping commitment strategy central — because teams cannot commit and central cannot right-size.

GCP Billing & Cost Optimization Interview Questions and Answers

How to use this set