CertifyCloudxAce every cloud certification.

PCA sample questions with answers

10 free practice questions for the Professional Cloud Architect exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Designing and planning a cloud solution architecture

Beckermet Chemicals keeps 600 TB of process history in Cloud Storage in europe-west2 and runs its modelling on a cluster in its own data centre, pulling about 80 TB a month across a Dedicated Interconnect. Egress is now the largest single line on the bill. The modelling code is already containerised, each monthly run reads the entire history from scratch, and the results it produces are around 200 GB. Which change reduces cost the most?

  1. A.

    Add a second Dedicated Interconnect circuit so that the monthly transfer completes in less elapsed time.

  2. B.

    Run the modelling containers in europe-west2 and transfer only the 200 GB of results out each month.

  3. C.

    Cache recently read objects on premises so repeated reads do not leave Google Cloud.

  4. D.

    Move the 600 TB to Coldline storage and keep reading it from the data centre.

Show answer

Answer: B

Compute is portable and 600 TB of data is not, so moving the containers next to the data cuts the egress from 80 TB a month to 200 GB.

  • A. Interconnect egress is charged per gigabyte transferred, so a second circuit adds fixed cost while moving the same volume.
  • B. Relocating portable compute next to immovable data replaces 80 TB of monthly egress with 200 GB, which is the dominant term on the bill.
  • C. A cache only helps when objects are read more than once, and the stem states each monthly run reads the entire history from scratch.
  • D. Coldline reduces storage price but raises retrieval charges, so data read in full every month costs more, and the egress is unaffected.
Question 2Designing and planning a cloud solution architecture

Loweswater Energy's continuity plan fails its trading platform over to a second region within 30 minutes. That region normally runs three instances; at failover an autoscaler must start 260 more of a memory-optimised machine type. A rehearsal at a quiet hour succeeded in 22 minutes. The risk committee has asked what happens if the primary region fails during a market event, when many other organisations would be moving into the same second region at the same moment. What should the plan add?

  1. A.

    Buy a three-year committed use discount covering the memory-optimised capacity in the second region.

  2. B.

    Spread the failover fleet across three zones of the second region using a regional managed instance group.

  3. C.

    Raise the autoscaler's maximum size in the second region and shorten its cool-down period.

  4. D.

    Reserve the failover capacity in the second region so those machines are held for this project.

Show answer

Answer: D

Only a reservation actually holds machines for you; everything else in the plan assumes capacity will be available on demand at the worst possible moment.

  • A. A committed use discount is a spending commitment for a lower price and carries no capacity guarantee whatsoever.
  • B. Spreading across zones protects against a zonal failure inside the recovery region but does nothing about regional capacity when everyone fails over at once.
  • C. Autoscaler limits govern how many instances are requested, not whether the region can satisfy the request when demand spikes across all tenants.
  • D. A reservation holds the specific machine shapes in the recovery region for this project, which is the only mechanism that guarantees capacity during a correlated regional event.
Question 3Designing and planning a cloud solution architecture

Oxenholme Insurance has signed a contract committing its new quote service to 99.95% availability measured monthly. The design calls an external pricing engine, operated by a third party under a 99.5% availability commitment, on every quote request before a price can be returned. The Google Cloud components are each designed to 99.99%. The third party will not improve its commitment, and the customer contract cannot be reopened. What should you change?

  1. A.

    Retry the pricing engine call with exponential backoff until it succeeds or the request deadline expires.

  2. B.

    Add a second Google Cloud region and raise the service's internal availability target to 99.999%.

  3. C.

    Return a cached indicative price immediately and settle the exact price asynchronously once the pricing engine responds.

  4. D.

    Deploy the quote service to three regions behind a global external Application Load Balancer.

Show answer

Answer: C

A dependency on the synchronous critical path caps the whole service at that dependency's availability, so the only design that can reach 99.95% is one that removes the pricing engine from the path a quote must traverse.

  • A. Backoff and retry absorb brief transient errors, not the multi-minute outages an availability commitment measures, and they turn failures into timeouts.
  • B. Raising an internal target that is not the binding constraint cannot move the composite figure above the weakest serial component.
  • C. Removing the third party from the synchronous path breaks the serial dependency, so its 99.5% commitment no longer bounds the composite availability of the quote service.
  • D. Multi-region deployment improves a term already designed to 99.99% while the 99.5% dependency remains in series and continues to cap the result.
Question 4Designing and planning a cloud solution architecture

Hellifield Telecom is designing a Spanner table for call detail records, expected to grow to 40,000 inserts per second. The proposed primary key is a monotonically increasing sequence number generated by the application, because the team wants records stored in arrival order for reporting. A reviewer warns that throughput will not scale with the number of nodes added. What should the schema use instead?

  1. A.

    The same sequence number, with additional Spanner nodes provisioned to absorb the concentrated write rate.

  2. B.

    A key beginning with the insertion timestamp, which spreads writes across the time dimension.

  3. C.

    A key whose leading component distributes writes evenly, such as a universally unique identifier or bit-reversed sequence.

  4. D.

    A key beginning with the originating switch identifier, of which the network has four in total.

Show answer

Answer: C

Spanner splits data by key range, so a monotonically increasing leading key sends every insert to the same split and no amount of extra capacity helps; the leading component must distribute.

  • A. Extra nodes cannot relieve a single hot split, because the constraint is that every insert belongs to the same key range.
  • B. A timestamp is monotonically increasing too, so a timestamp-leading key concentrates writes at the end of the key space exactly as the sequence does.
  • C. A uniformly distributed leading key spreads concurrent inserts across many splits, which is what lets write throughput grow with instance size.
  • D. Four distinct leading values create only four hot ranges, so parallelism is capped far below what the write rate requires.
Question 5Designing and planning a cloud solution architecture

Sedbusk Genomics trains models on 900 TB of imagery held in Cloud Storage. Training runs read the whole dataset repeatedly across 500 accelerator nodes and are currently accelerator-bound at 30% utilisation because data arrives too slowly. The dataset is written once by an upstream pipeline and never modified during training. Cost matters, but idle accelerators cost more than storage. Which two design choices should you make? (Choose TWO.)

Choose 2.

  1. A.

    Keep Cloud Storage as the durable system of record and hydrate the file system from it before each campaign.

  2. B.

    Give each accelerator node its own persistent disk copy of the full 900 TB dataset.

  3. C.

    Stage the training data into a high-throughput parallel file system such as Managed Lustre for the run.

  4. D.

    Move the dataset permanently to a regional Filestore instance and delete the Cloud Storage copy.

  5. E.

    Convert the dataset to the Coldline storage class to reduce the cost of holding 900 TB.

  6. F.

    Reduce the number of accelerator nodes so that the available read throughput keeps them busy.

Show answer

Answer: A, C

Object storage is the right durable home for 900 TB but not the right read path for 500 nodes re-reading it constantly; stage it into a parallel file system for the run and keep the bucket as the system of record.

  • A. Keeping the bucket as the system of record preserves durability and cheap retention while the file system exists only for the duration of a campaign.
  • B. Replicating 900 TB to each of 500 nodes is hundreds of petabytes of duplication and is impossible within any sane budget or provisioning time.
  • C. A parallel file system delivers the aggregate throughput and POSIX semantics that hundreds of training clients re-reading one dataset require.
  • D. Deleting the object copy trades away durability and low-cost long-term retention, and a general file service is not built for this aggregate read rate.
  • E. Coldline charges for retrieval on data that is read in full repeatedly, so it increases both cost and read latency.
  • F. Shrinking the fleet accommodates the bottleneck rather than removing it, lengthening every training run at the same total accelerator cost.
Question 6Designing and planning a cloud solution architecture

Warcop Health's platform design treats every dataset identically: one project, one encryption key and one set of IAM grants. The estate holds published tariff data, internal operational data and clinical records. A reviewer objects that the design cannot demonstrate that clinical records are handled differently from tariffs. Applying the strictest controls everywhere would make routine analytical work unworkably slow. What should the design introduce?

  1. A.

    The strictest available controls applied uniformly to every dataset, so that nothing can be under-protected anywhere.

  2. B.

    A quarterly access review in which dataset owners confirm who is currently able to read each dataset.

  3. C.

    A data classification scheme whose levels drive placement, key management and access for each dataset.

  4. D.

    Sensitive Data Protection scans reporting where clinical identifiers appear across the existing estate.

Show answer

Answer: C

Proportionate protection requires the design to distinguish categories of data first; a classification scheme is what lets placement, keys and access differ by sensitivity instead of being uniform.

  • A. Uniform maximum controls breach the stated constraint by making ordinary analytical work on public tariff data unworkably slow.
  • B. Periodic review detects inappropriate access after it has been granted rather than shaping how the design grants it.
  • C. Classification levels give the design a basis for differentiating placement, keys and access, and new datasets inherit the right treatment automatically.
  • D. Discovery scanning finds where sensitive identifiers occur, which informs classification but does not itself differentiate how datasets are protected.
Question 7Designing and planning a cloud solution architecture

Wharfedale Archives must move a 190 TB object store into Cloud Storage. The copy will take about eleven days over the available link. The archive is not frozen: roughly 900 GB of new objects are added and a few thousand existing objects are amended each day. The cutover to the new location must not lose an amendment, and the business will accept an application freeze of no more than two hours. How should the plan sequence the move?

  1. A.

    Have the application write to both locations for eleven days and copy the historical objects afterwards.

  2. B.

    Copy everything once and accept that objects changed during the eleven days will be corrected later.

  3. C.

    Run a bulk copy, then repeated incremental syncs of changes, then a final sync inside the freeze.

  4. D.

    Freeze the archive for eleven days and copy everything once, so no changes occur.

Show answer

Answer: C

Bulk copy, converge with incremental syncs, then cut over during a short freeze: the standard sequence for moving a dataset that keeps changing while it is being copied.

  • A. Dual writes require application changes and partial-failure handling, and the historical objects still need the same bulk-plus-delta treatment afterwards.
  • B. Nothing in this approach identifies which objects changed during the copy, so amendments are silently lost — the one outcome the stem forbids.
  • C. Successive incremental passes shrink the delta until it fits inside the agreed freeze, which is what makes a short cutover possible on a long copy.
  • D. An eleven-day freeze massively exceeds the two-hour outage the business has agreed to accept.
Question 8Designing and planning a cloud solution architecture

Ravenstone Auctions is splitting a monolith into a bidding service and an inventory service. A bid may only be accepted if the lot is still available, and regulation requires that an accepted bid and the reservation of the lot are either both recorded or neither is. The team proposes that bidding publish an event and inventory consume it asynchronously, arguing that decoupling is a Google-recommended practice. What should you advise?

  1. A.

    Adopt the asynchronous event, because decoupled services are more available and scale independently.

  2. B.

    Keep bidding and inventory in one service with a single database transaction covering both records.

  3. C.

    Publish the event asynchronously and add a nightly reconciliation job that repairs any mismatched records.

  4. D.

    Have bidding call inventory synchronously and retry the call until inventory confirms the reservation.

Show answer

Answer: B

A regulatory all-or-nothing requirement across two records is a transactional boundary, and a service boundary drawn through a transaction creates a consistency problem no amount of messaging removes.

  • A. It applies a general principle to a case whose stated regulatory constraint the principle cannot satisfy, leaving a window where one record exists without the other.
  • B. A single transaction is the only mechanism that satisfies an all-or-nothing requirement simply, and it recognises that the two records form one consistency domain.
  • C. Reconciliation repairs inconsistency after it has occurred, whereas the requirement is that the inconsistent state must never be recorded at all.
  • D. Synchronous retries still leave a failure window where the bid is recorded and the reservation is not, while giving up the independence that motivated splitting.
Question 9Designing and planning a cloud solution architecture

Ravenstonedale Media transcodes video in a Cloud Run function invoked over HTTP by its upload service. Short clips succeed. Longer files return 504 to the caller once the configured timeout elapses, the invocation is billed for its full duration, and the upload service's retry starts the transcode again from the beginning. Transcodes take between four minutes and two hours and need several cores and a large scratch area. What should you do?

  1. A.

    Add a retry policy with exponential backoff so that a failed transcode eventually completes successfully.

  2. B.

    Raise the function's timeout and memory allocation to the maximum values the platform permits.

  3. C.

    Split each file into 60-second chunks so that every invocation finishes well inside the timeout.

  4. D.

    Accept the upload, return a job identifier immediately, and run the transcode as a Cloud Run job or a Batch task.

Show answer

Answer: D

The work is a long-running batch task wearing a synchronous request's clothes; the fix is to decouple acceptance from execution and run the transcode on a platform built for tasks rather than requests.

  • A. Retrying a request that cannot finish inside the available time repeats the whole transcode and multiplies the wasted cost.
  • B. It raises a ceiling rather than removing the synchronous model, and a two-hour transcode will exceed the higher ceiling as well.
  • C. Chunking adds segmentation and reassembly complexity while leaving the caller waiting synchronously for the complete result.
  • D. Returning immediately removes the request timeout entirely and a task platform supplies the duration, CPU and scratch space a long transcode needs.
Question 10Designing and planning a cloud solution architecture

Hartsholme Rentals' engineering lead proposes a multi-region Spanner instance for a new booking database. The service operates in one country, all writes originate there, the agreed availability target is 99.95%, peak load is 400 writes per second, and the finance director has asked why the database line is four times the estimate for the rest of the platform. The lead argues that multi-region avoids a rebuild if the company expands abroad. Which design decision should you take?

  1. A.

    Use Firestore in Native mode, which scales automatically and has no instance configuration to choose.

  2. B.

    Use a regional Spanner instance now, since a regional configuration already exceeds the agreed availability target.

  3. C.

    Keep the multi-region instance, because reconfiguring the database later on would require an expensive and risky migration.

  4. D.

    Use Cloud SQL for PostgreSQL with a read replica, since 400 writes per second is modest for one instance.

Show answer

Answer: B

A regional Spanner configuration already meets the stated availability target for a single-country, single-write-region workload, so the multi-region premium buys a requirement nobody has asked for.

  • A. A document store with a different consistency and query model is a far larger change than the cost question warrants, and the booking workload is relational.
  • B. A regional configuration already exceeds the agreed availability target for a workload whose writes all originate in one region, so the multi-region premium is unfunded by any requirement.
  • C. It pays continuously for a capability with no current requirement, and a later configuration change is a data migration rather than the rebuild claimed.
  • D. It changes the database engine on the basis of today's peak, giving up the horizontal write scaling that motivated choosing Spanner in the first place.

Keep going with 521 more PCA questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all PCA papers
PCA sample questions with answers (10 free) · CertifyCloudx