aws · google cloud · azure · one job, three names

The three-cloud atlas: same jobs, three dictionaries

Three companies sell roughly the same shop, and each wrote its own dictionary. A bottomless bucket for files is S3 in one, Cloud Storage in the next and Blob Storage in the third. This lab is the whole dictionary: every service each provider lists today, mapped onto the jobs those services actually do, with the small print that makes “the same thing” not quite the same. Type any name you know and read across. Then let the atlas tell you what has been renamed or retired since you last looked, because a cloud vocabulary from two years ago is already wrong in places.

Two neighbours. The Cloud lab explains how the machinery works: requests, networks, identity. Its Cloud differences tab holds the handful of structural differences that change an architecture. This atlas is the names: what each provider sells, what it calls it, and what it costs you to choose it.

Three shops, one stock list: the names differ, the small print differs more

In plain English: underneath the branding, the three big clouds rent out the same dozen kinds of thing: machines, storage, databases, networks, queues, identity, monitoring, and lately models and agents. So the first job is translation, and it is mostly mechanical: learn one cloud well and the second is a dictionary of new names for old ideas. The second job is the one people skip. Two services with the same job are rarely the same shape. One bucket can span a continent; another cannot. One archive returns a file in milliseconds; another makes you wait hours. Those differences are what this atlas is for, and every row carries them.

How to read a row

Try this: every job in the atlas is one row with the same seven parts. Read them in this order and a row takes twenty seconds.
the jobwhat you are trying to do, in words that do not belong to any cloud: “object storage” (keeping files by name, any size, for as long as you like), not S3.
three nameswhat each provider calls its answer. A chip beside the name says whether it is GA (generally available, meaning sold and supported), in preview (usable, not promised), retiring (closed to new customers or given an end date) or retired.
no twinhonest gaps. Where a provider has nothing under that heading, the row says so instead of inventing a match.
buys you / bills youthe pros and cons of each provider’s version, written as what you gain and what you pay for it, in time, money or freedom.
pick whenthe situation in which one of the three is the obvious choice, if there is one.
not the same thingthe cross-cloud trap: where “equivalent” services behave differently in a way that changes a design.
words on this rowevery acronym the row uses, defined right there, so you never leave the row to look one up.

Try it: name that cloud

Try this: ten real service names with the brand word removed. Guess which cloud sells each one. You will be wrong a few times. The pattern in the wrong ones is the lesson: each provider has a naming habit, and once you hear it you can place a service you have never met.
in real lifeThree supermarkets, same aisles, different own-brand labels. After a few visits you can tell which shop a tin came from by the typeface alone. Cloud names work the same way: one prefers a plain noun, one prefixes its company name, one names things after what they do.
start byread the first name aloud, guess, and read why under the answer before moving on; the reasons add up to three naming habits.
words hereAWS Amazon Web Services, Amazon’s cloudAzure Microsoft’s cloudGoogle Cloud Google’s cloud, often written GCP for Google Cloud Platformservice one thing a cloud sells you as a unit: a kind of storage, a kind of database, a toolbrand word the company part of a name, such as Amazon, AWS, Azure, Microsoft or Google, removed here so the name itself has to do the workGA generally available: sold and supported, not a preview

Try it: the atlas, every job across three clouds, with the small print

Try this: search by any name you already know, from any cloud, and read across. Click a row to open it: three cards, one per provider, each with what it buys you and what it bills you, then when to pick which, and the trap. Choose your home cloud and the atlas puts your cloud first on every row, so the other two read as translations of the one you know.
in real lifeA phrasebook laid out as a table: one column per language, one row per thing you might need to say. The useful phrasebooks add a footnote when the literal translation will get you into trouble, and that footnote is the column on the right.
start bytype queue, open the row, and read the “not the same thing” box: one provider sells two products where the others sell one.
words herejob the cloud-neutral thing you need done, the unit of a rowhome cloud the provider you know best; the atlas puts it first and phrases the others as translationscore / common / niche how often teams need the job: almost every system, many systems, or a specialist fewGA generally available: sold and supportedpreview usable now, not yet promised; features and prices can changeretiring closed to new customers or given an end date; do not start new work on itretired no longer offeredno twin the provider has no product under this headingbuys you / bills you the pros and cons: what you gain, and what you pay for it in money, time or freedommanaged the provider runs the servers and upgrades; you use the serviceregion one geographic area of data centres; a zone is one isolated site inside it
home cloud
how common

Try it: every service, one provider at a time

Try this: the complete catalogue of one provider, grouped by family, with its status and its former name where it was renamed. Each entry that has twins in the other two clouds says so, and clicking the twin opens that row in the atlas. Filter to retiring and you have the list of things not to start new work on.
in real lifeThe full product catalogue of one shop, shelf by shelf, with a sticker on each item: in stock, new and untested, or being discontinued. The atlas tab is the comparison; this tab is the inventory.
start bypick Azure, filter to retiring, and count how many of them you have seen in an architecture diagram this year.
words herecatalogue the provider’s own list of everything it sells, as of this atlas’s datefamily one of fourteen groupings used across this lab, such as Storage or Databases, so the three catalogues line upGA generally available: sold and supportedpreview usable now, not yet promisedretiring closed to new customers or given an end dateretired no longer offeredwas the previous name of a renamed servicetwins the services in the other two clouds that do the same job
status

Try it: the depth map, where each cloud has the most on the shelf

Try this: a heat map counted live from the catalogue: how many services each provider lists in each family. Darker means more. Click a cell to see exactly which services are being counted. Then switch to GA only and watch some of the depth evaporate, because a thick shelf of previews is not the same as a thick shelf.
in real lifeThree hardware shops. One has an entire aisle of screws; another has a shelf. That tells you where each shop concentrates, not which screw is better. Breadth is a clue about focus, not a verdict on quality.
start byfind the family where the three counts differ most, click the smallest cell, and ask what that provider expects you to do instead.
words herefamily one of fourteen groupings used across this lab, such as Storage or Databasesheat map a grid where a darker cell means a bigger numberGA generally available: sold and supportedpreview usable now, not yet promisedretiring closed to new customers or given an end datebreadth how many distinct services a provider lists; not the same as how good any one of them is
count

Try it: build a stack, then swap the cloud underneath it

Try this: pick a workload and the atlas assembles the services it needs, layer by layer, in one cloud. Press swap cloud and every box relabels itself in place with the equivalent from the next provider. Below the diagram is the bill of materials: one thing each chosen service buys you and one thing it bills you, straight from the atlas rows. The point is to feel how little changes in shape, and how much changes in the small print.
in real lifeRebuilding the same kitchen in a different country. The layout is identical: sink, hob, fridge, extractor. Every appliance has a different brand, a different plug, and one of them turns out not to exist there at all.
start bychoose AI agent platform, press swap cloud twice, and read the bill of materials each time: the boxes keep their places, the cons change completely.
words hereworkload one kind of system you might build, such as a web app or a data platformstack the set of services a system is built from, drawn as layerslayer one band of the stack: the front door, the code, the data, the glue, the watchingbill of materials the list of parts used, here the services chosen and one pro and one con of eachno twin the provider has nothing for that box; the diagram leaves it hollow rather than guessingGA generally available: sold and supported
workload
cloud

What changed: launches, renames and retirements since 2024

In plain English: cloud vocabularies rot. A service you learned three years ago may now have a new name, a new owner inside the company, or an end date. This timeline is the change log the providers do not publish in one place, sorted by date: announced future retirements first, then the newest changes, filterable by provider and by kind of change. If a name in your notes is not here and not in the atlas, it is worth checking before you say it in an interview.
provider
kind

Try it: which name is current?

Try this: eight quick rounds drawn from the change log. Sometimes two names for the same thing, one current and one old; sometimes a service, and the question is whether it is still sold; sometimes a recent launch and the question is whose it is. Each answer explains the change in one line.
in real lifeCalling a company by the name it had before the merger. Everyone knows what you mean; everyone also knows how long ago you last looked.
start byanswer from memory first, then read the one-line reason; the reason is the part that sticks.
words herecurrent name what the provider calls the service todayformer name what it was called before a renameretired no longer sold; existing customers may have a deadline to leaveretiring closed to new customers or given an end date that has not yet arrivedlaunched new since 2024, sometimes still in previewpreview usable now, not yet promised

Try it: the scenario drill

Try this: the scenarios below, shuffled, with the wrong answers taken from other scenarios on this page so they are plausible by construction.
in real lifeThis drill is the flashcard: a situation, and you pick the right service from look-alikes borrowed from the other scenarios. The list underneath is the revision guide. Read it after, not before, or you are only recognising, not producing.
start bypress next scenario before reading the list below.
words herescenario a situation described the way an exam or an interviewer would put itdistractor a wrong option that is a real answer to something else

When the scenario says… which service, in which cloud

In plain English: interviewers and exams rarely ask “what is Spanner”. They describe a constraint and wait to see which name you reach for. Each line below is a constraint, then the answer in all three clouds, then the reason the other options lose.
“We need one relational database with strong consistency that scales writes across regions.” → Spanner on Google Cloud is the reference answer, and Aurora DSQL is AWS’s first true equivalent since 2025. Aurora Global Database is read replicas plus a promotion, not multi-region writes, and Azure has no relational equivalent: Cosmos DB offers multi-region writes but it is not a relational database.
“A serverless warehouse where we pay only for what each query scans.” → BigQuery. On AWS the closest are Redshift Serverless and Athena, and on Azure the Fabric warehouse or Synapse serverless SQL. Redshift provisioned clusters and dedicated SQL pools are the opposite model: you size and pay for capacity whether or not you query.
“Our Kafka producers must keep working with no code change.” → Azure Event Hubs speaks the Kafka protocol natively, so clients repoint with a connection string. On AWS use Amazon MSK, on Google Cloud Managed Service for Apache Kafka; Kinesis and Pub/Sub are their own protocols and need client changes.
“Managed Kubernetes with the least node management we can get away with.” → GKE Autopilot has done it longest; EKS Auto Mode and AKS Automatic are the 2024 and 2025 answers on the other two. Standard EKS and AKS still leave node pools, upgrades and capacity to you.
“We already pay for Windows Server and SQL Server licences.” → Azure Hybrid Benefit lets those licences carry into Azure VMs and Azure SQL, and it frequently decides the whole business case. AWS and Google Cloud offer bring-your-own-licence paths for some products, but nothing as broad or as cheap.
“Reach a partner’s service privately, and our address ranges overlap with theirs.” → A private endpoint: AWS PrivateLink, Google Cloud Private Service Connect, Azure Private Link. The overlap is irrelevant because the two networks never join; peering or a VPN would fail on exactly that overlap.
“Guardrails that can also fix a non-compliant resource, not only deny the request.” → Azure Policy, whose effects include audit, deny and deploy-if-not-exists remediation. AWS service control policies only deny, and Google Cloud Organization Policy constrains what may be created; both need a separate remediation service to change anything.
“Stop data leaving our project even when the identity is authorised.” → VPC Service Controls on Google Cloud: a perimeter around services that blocks calls from outside it regardless of IAM. Azure’s Network Security Perimeter is the closest packaged equivalent; AWS assembles the effect from VPC endpoint policies and service control policies rather than selling it as one product.
“Run an agent in production with managed sessions, memory and a gateway to its tools.” → Amazon Bedrock AgentCore, Google Cloud’s Agent Runtime with Agent Gateway inside the Gemini Enterprise Agent Platform, and Azure AI Foundry Agent Service. Bedrock, Vertex AI and Foundry Models alone give you model access, not the operating layer around the agent.
“Screen prompts for injection and responses for leaks before they reach the model or the user.” → Amazon Bedrock Guardrails, Google Cloud Model Armor, Azure AI Content Safety. A web application firewall is the wrong layer: it inspects HTTP traffic, not the meaning of a prompt.
“Where did Azure Active Directory go?” → It became Microsoft Entra ID in 2023 and the Entra family now covers workforce identity, external identity and permissions. Same service, same tenants; only the name and the product grouping changed.
“Which provider should host our source code?” → Usually GitHub or GitLab, with the cloud’s own build service behind it. AWS closed CodeCommit to new customers in 2024 and quietly reopened it in late 2025, Google Cloud is retiring Cloud Source Repositories in favour of Secure Source Manager, and Azure points you at Azure Repos or GitHub, which Microsoft owns. None of the three treats hosting your code as a flagship.
“Archive petabytes cheaply, but a restore must return a file in milliseconds.” → Google Cloud Storage Archive class keeps millisecond access at archive prices, and S3 Glacier Instant Retrieval does the same on AWS. Azure’s Blob Archive tier and S3 Glacier Deep Archive need rehydration measured in hours, which changes a disaster-recovery design.
“Redis changed its licence. What do the clouds sell now?” → ElastiCache and Memorystore both added Valkey, the open-source fork, and steer new work to it; Azure Managed Redis is built on Redis Enterprise under a partnership. The old “managed Redis” row is now three different answers.
“A container orchestrator that is not Kubernetes.” → Amazon ECS, in practice. Azure Service Fabric exists but is legacy-leaning, with Microsoft steering new work to AKS and Container Apps, and Google Cloud has none: container work there goes to GKE or Cloud Run.