Choosing Storage & Databases

Google ACE ACE · Planning and implementing a cloud solution

By Abu Anns · Last updated 20 November 2026Exam version: ACE (exam guide live 30 June 2026)

Two kinds of storage decisions

Storage planning on Google Cloud involves two different decisions, and the exam tests both. The first is object and block storage: where do files and VM disks live, and how often are they accessed? The second is database choice: which managed database fits the data model and access pattern? The two decisions are independent — a single application might store user uploads in Cloud Storage, its transactional data in Cloud SQL, and its analytics in BigQuery — so the skill the exam tests is matching each data need to the right service, not finding one service for everything.

Cloud Storage classes

Cloud Storage is Google's object storage: durable, highly available storage for files of any kind, accessed over HTTP. Buckets have a location (regional, dual-region, or multi-region) and a storage class that trades access cost against storage cost. The classes form a spectrum by access frequency: Standard for frequently accessed data (hot), Nearline for data accessed about once a month, Coldline for data accessed about once a quarter, and Archive for data accessed less than once a year. The colder the class, the cheaper the storage and the more expensive each retrieval — and the longer the minimum storage duration (30 days for Nearline, 90 for Coldline, 365 for Archive), with early-deletion fees if you remove data before it. Choosing the wrong class is a classic cost mistake: Archive storage for data you read weekly costs more in retrieval fees than it saves on storage.

A bucket's class can be changed, and Object Lifecycle Management can automate the transition: rules that move objects to a colder class after a number of days, or delete them after a retention period. Autoclass goes further and moves objects between classes automatically based on observed access patterns. The exam also tests bucket location choices: a regional bucket keeps data in one region (cheapest, and required for data-residency rules), while multi-region buckets (such as EU or US) replicate across regions for higher availability. Dual-region buckets replicate between two specific regions. A bucket's location cannot be changed after creation — you must create a new bucket and copy the data.

Persistent Disk

Persistent Disk is block storage for Compute Engine VMs and GKE nodes: the disks that VMs boot from and store data on. Disks come in two durability scopes: zonal disks live in a single zone, and regional disks synchronously replicate between two zones in the same region, surviving a zone failure. Within each scope there are performance tiers: standard persistent disk (HDD, cheapest, for throughput workloads), balanced persistent disk (SSD, the general-purpose default), and SSD persistent disk (highest IOPS, for latency-sensitive databases). Exam questions often hinge on the trade-off: a regional SSD disk is the answer when the requirement is both high availability and high performance, while a zonal standard disk is the answer for a cost-sensitive batch VM. Disks can be resized while attached, and snapshots provide point-in-time backups.

Choosing a database

Google Cloud's managed databases each fit a different shape of data, and the exam tests the mapping:

ServiceTypeBest for
Cloud SQLRelational (MySQL, PostgreSQL, SQL Server)Standard relational workloads, lift-and-shift of existing databases
AlloyDBRelational (PostgreSQL-compatible)Demanding PostgreSQL workloads needing higher performance and availability
SpannerRelational, globally distributedGlobal scale with strong consistency and horizontal scaling
FirestoreDocument (NoSQL)Mobile and web apps, flexible documents, real-time sync
BigtableWide-column (NoSQL)Massive-scale analytical and time-series workloads, millions of operations per second
BigQueryData warehouse (analytics)Analyzing terabytes to petabytes with SQL, not transactional workloads

The most common exam confusions are worth naming. BigQuery is not a transactional database: it is an analytics warehouse, and using it as the backing store for an application's day-to-day reads and writes is wrong. Firestore is not for analytics: it is a document database for application data. Spanner is for global relational scale: if the requirement mentions multiple regions with strong consistency and relational data, Spanner is the answer; if it mentions a single region and a standard relational workload, Cloud SQL is cheaper and simpler. Bigtable is for enormous scale: time-series data, IoT telemetry, or workloads needing millions of reads and writes per second — it is overkill for an ordinary application database.

Planning questions: how the exam asks

Planning questions usually describe a workload in a sentence or two and ask for the best storage or database choice. The signal words matter: "frequently accessed images" points to Standard storage; "backups accessed rarely" points to Coldline or Archive; "relational data, single region" points to Cloud SQL; "petabytes of analytics with SQL" points to BigQuery; "millions of sensor readings per second" points to Bigtable; "global users, strong consistency" points to Spanner. When two services seem plausible, look for the detail that separates them — usually scale, consistency, or access frequency — and choose the simplest service that meets the stated requirement. The exam rewards the least complex adequate answer, not the most powerful service.

Durability, availability, and cost levers

Cloud Storage is designed for 99.999999999% (eleven nines) annual durability — objects are stored redundantly across multiple devices, and in multi-region and dual-region buckets, across multiple regions. That durability is the same across storage classes: Archive data is just as durable as Standard data; the classes differ in access cost and retrieval speed, not in how safely the bytes are kept. Availability differs by location type: multi-region buckets offer the highest availability SLA (99.95%), regional buckets slightly less (99.9%). For the exam, the practical point is that choosing a colder class never trades away durability — it trades access cost and minimum storage duration for cheaper storage, which is why lifecycle policies can safely move aging data to colder classes.

Cost planning for storage has three levers, and the exam tests all of them. The first is storage class: match the class to the access pattern, and use lifecycle rules or Autoclass to move data automatically as it ages. The second is location: regional storage is cheaper than multi-region, so data that does not need geographic redundancy should stay regional — and data-residency requirements may force a regional choice anyway. The third is egress: moving data out of Google Cloud (to the internet or to another cloud) is charged, while ingress is free, so an architecture that constantly downloads large datasets will cost more than one that processes data where it sits. A common exam trap is a design that stores data in one region and processes it in another: cross-region traffic adds latency and cost, so the storage and the compute should usually live in the same region.

For databases, the cost levers are instance size, storage, and backups. Cloud SQL charges for the machine type, the storage provisioned, and the backups retained — right-sizing the instance and enabling automatic storage increases (rather than over-provisioning upfront) keeps costs down. BigQuery charges primarily for the data scanned by queries, so partitioning and clustering tables is the main cost control. Firestore charges per document read, write, and delete, so chatty access patterns cost more than batched ones. Knowing which lever belongs to which service is what turns a generic "reduce costs" question into a specific, correct answer.

Advertisement

Exam traps to watch for

Practice questions

Test yourself on this guide. Answers are checked instantly, with a full explanation for each question.

Q1.A company stores regulatory records that must be kept for seven years and are almost never accessed. Which Cloud Storage class is the most cost-effective choice?

Report an error

Q2.A mobile app needs a flexible NoSQL database that stores user profiles as documents and syncs changes to devices in real time. Which service is the best fit?

Report an error

Q3.A database VM runs a latency-sensitive application and must keep working if its zone fails. Which Persistent Disk configuration should you choose?

Report an error

Q4.A company needs to analyze 50 TB of historical sales data with SQL queries, run a few times a day by analysts. The data does not change during the day. Which service is the best fit?

Report an error

Q5.An application serves user-uploaded photos to a global audience, and the photos are accessed frequently. The bucket is currently in a single region and users in other continents report slow loads. What is the best change?

Report an error

Q6.A financial services company needs a relational database that serves users in North America, Europe, and Asia with strong consistency and horizontal scaling. Which service is the best fit?

Report an error

Q7.An IoT platform ingests millions of sensor readings per second and needs to store them for time-series analysis. Which service is the best fit?

Report an error

Q8.A company is migrating an existing on-premises MySQL database to Google Cloud with minimal changes to the application. Which service is the best fit?

Report an error