Storage

Three kinds of storage for three ways of reaching data

Object storageBucketwhole files, fetched by nameover the webS3 · Blob Storage · Cloud StorageBlock storageone machineits diska raw disk, attached toone machineEBS · Managed Disks · Persistent DiskFile storagea shared folder, opened bymany machines at onceEFS · Azure Files · FilestoreThree kinds of storage, told apart by one question: how is the data reached?By name over the web, as a disk inside one machine, or as a folder shared across a network. Everything else, price,speed, limits, follows from that. Most systems use all three: a disk for the database, a bucket for the photos and backups,and a shared folder only when old software insists on one.

Object, block and file storage differ in how data is reached: by name over the web, as a disk, or as a shared folder.

Object storage

Whole files under names, with no limit on how many

Bucket: shop-mediaa flat list, any lengthphotos/2025/cat.jpg2.1 MBphotos/2025/dog.jpg1.8 MBbackups/db-0612.gz40 GBlogs/2025-06-12.txt3 KBOne objectkeyphotos/2025/cat.jpgdata2.1 MB of bytes, stored wholemetadatatype image/jpeg, date, your own tagsversionolder copies kept, if you switch it onhttps://shop-media.s3.amazonaws.com/photos/2025/cat.jpgevery object has a web addressObject storage keeps whole files under names, in a bucket with no size limit.There are no real folders: the slashes are part of the name. You put, get and delete whole objects through a web API;you cannot change the middle of one. In return it scales without planning, costs the least, and anything on the internetthat you allow can read it. Photos, video, backups, logs and data for analysis all live here.

An object is data plus a key plus metadata, kept in a bucket and reached through a web address.

Object storage · Going deeper

Buckets, containers and the switches that matter

AWSAzureGoogle CloudThe serviceS3Blob StorageCloud StorageThe container for objectsbucketcontainer, in a storage accountbucketOne stored itemobjectblobobjectKeep old versionsversioningblob versioningobject versioningCopy to another regioncross-region replicationgeo-redundant storagedual-region, multi-regionA bucket on AWS and Google Cloud is a container inside a storage account on Azure.Bucket names on S3 are unique across all AWS customers, because each becomes part of a web address. New buckets areprivate and encrypted by default on all three providers. Four switches matter in practice: versioning, to recover fromoverwrites and deletions; replication, to survive a region; lifecycle rules, to cut cost; and public access, left off.

S3 buckets, Azure containers in a storage account and Cloud Storage buckets do the same job. Versioning, replication, lifecycle rules and blocked public access are the settings to know.

Durability and availability

Eleven nines: every object is copied across three zones

you storeone objectZone aZone bZone ccopies on several disks in each of three zones99.999999999%durability, called eleven nines:store 10 million objects, and expect tolose one of them every 10,000 years.99.99%availability, a different promise:reachable all but about 53 minutes a year.Durability is whether the data still exists. Availability is whether you can reach it right now.Object storage is built to be almost impossible to lose: the write is confirmed only after copies are safe in three zones,and damaged copies are repaired from the others. An outage makes data unreachable for a while; it does not destroy it.Durability does not protect you from yourself: delete or overwrite an object and all the copies obey. That is what versioning is for.

Durability is the chance data survives; availability is the chance you can reach it now. Object storage offers 11 nines of the first.

Durability and availability · Going deeper

Azure’s four redundancy options, drawn as copies

LRSlocally redundantone data centre11 ninesdurability a yearZRSzone-redundantzone 1zone 2zone 312 ninesdurability a yearGRSgeo-redundantprimary regionpaired region16 ninesdurability a yearGZRSgeo-zone-redundantzone 1zone 2zone 3paired region16 ninesdurability a yearAzure makes the copies a choice: three in one building, across three zones, or also in a second region.Each purple bar is one copy. LRS survives a failed disk or rack; ZRS survives a lost data centre; GRS and GZRS survivea lost region, with the far copies written a little later. More copies cost more. This table is asked directly on AZ-900.AWS S3 and Google Cloud Storage spread across zones by default, and offer single-zone and multi-region options beside it.

LRS keeps three copies in one data centre, ZRS one in each of three zones, GRS and GZRS add three more in a paired region.

Storage classes

Pay less to store, and more to read

to store 1 GB for a monthto get it backminimum stayStandard$0.023millisecondsno minimumStandard-IA (infrequent access)$0.0125milliseconds, plus a fee per GB read30 daysGlacier Instant Retrieval$0.004milliseconds, higher fee per GB read90 daysGlacier Flexible Retrieval$0.0036minutes to 12 hours90 daysGlacier Deep Archive$0.00099within 12 hours180 daysColder classes store each gigabyte for less, and charge more, in money or waiting, to read it.All five keep the same eleven nines of durability. You pick by how often the data will be read: daily, monthly, oncea year, or probably never. Deep Archive costs one twenty-third of Standard. Bars are to scale; AWS S3 list prices,N. Virginia, 2025. Leaving a class before its minimum stay is charged as if the object had stayed.

A storage class trades a lower price per gigabyte for slower or costlier retrieval. Pick by how often data is read.

Storage classes · Going deeper · 1 of 2

The same ladder on Azure and Google Cloud

How often it is readAWS S3Azure BlobGoogle Cloud StorageEvery dayStandardHotStandardAbout once a monthStandard-IACool (30 days)Nearline (30 days)A few times a yearGlacier Instant RetrievalCold (90 days)Coldline (90 days)Almost neverGlacier Flexible, Deep ArchiveArchive (180 days)Archive (365 days)Let the provider decideIntelligent-Tieringlifecycle rulesAutoclassEvery provider has the same ladder: hot, cool, cold, archive.Days in brackets are the minimum stay. Azure’s Archive tier is offline: a blob must be “rehydrated” to another tier,which takes hours, before it can be read. Google’s Archive answers in milliseconds but has the longest minimum stay.When access is unpredictable, S3 Intelligent-Tiering and Google Autoclass watch each object and move it for you.

Hot, Cool, Cold and Archive on Azure; Standard, Nearline, Coldline and Archive on Google Cloud. Intelligent-Tiering and Autoclass choose for you.

Storage classes · Going deeper · 2 of 2

Lifecycle rules: data cools as it ages

One rule, written once: what happens to a 1,000 GB backup as it agesStandard$23 a monthday 0Standard-IA$12.5 a monthday 30Glacier Flexible$3.6 a monthday 90Deep Archive$0.99 a monthday 365deletedafter 7 yearsBar height is the monthly cost. The time axis is not to scale, so the early steps are visible.A lifecycle rule moves data to colder classes as it ages, and deletes it when its time is up.Data is read most when it is new. Without a rule, this backup would cost $23 a month for seven years: about $1,930.With it, the same seven years cost about $150. The rule is a few lines of configuration on the bucket, and it is thesingle most effective cost control in storage. Computed from the prices on the previous page.

A lifecycle rule moves objects to colder classes on a schedule and deletes them at the end. Here it cuts seven years of storage from about $1,930 to about $150.

Block storage

A disk attached to one machine

Zone aVirtual machineattachedA 100 GB volume30 GB used100 GByou pay for all 100 GB:the size you asked forSnapshota point-in-time copy,kept in object storageback up, restore, or copy to another zone or regionBlock storage is a virtual disk: fixed size, attached to one machine in the same zone.The operating system sees an empty disk, formats it and reads or writes any small block of it in under a millisecond.That makes it right for boot disks and databases, where files change in place. It is billed for the size you provisioned,used or not, and it outlives the machine: stop the machine and the disk, and its bill, remain.

A volume is a virtual disk of fixed size for one machine in the same zone, backed up with snapshots.

Block storage · Going deeper · 1 of 2

Kinds of disk: general purpose, fast, cheap and temporary

Kind of diskForAWS EBSAzureGoogle CloudGeneral purpose SSDboot disks, most workgp3Premium SSDBalanced PDHigh-performance SSDlarge, busy databasesio2Ultra DiskExtreme PDHard disk (HDD)big files read in orderst1, sc1Standard HDDStandard PDLocal, temporaryscratch space, cachesinstance storetemporary diskLocal SSDDisks are sold by speed. Speed is measured in IOPS: read or write operations per second.A general purpose SSD suits nearly everything: AWS gp3 gives 3,000 IOPS at $0.08 per GB a month. High-performance diskslet you buy tens of thousands of IOPS for a busy database. Hard disks are cheaper per GB and slow at random access.Local disks sit inside the physical server: fastest of all, and wiped when the machine stops. Never keep the only copyof anything on one.

Disks are sold by IOPS. General purpose SSDs suit most work; local disks are fastest and are wiped when the machine stops.

Block storage · Going deeper · 2 of 2

Snapshots copy only what changed

Monday: first snapshotcopies every block: 100 GBTuesdaycopies the 4 blocks that changed: 10 GBWednesdaycopies 2 changed blocks: 5 GBStored and billed: 115 GB for three days of history, not 300 GB. Any of the three days can be restored in full.Snapshots are incremental: after the first, each one stores only the blocks that changed.A snapshot is a backup of a whole disk at one moment, kept in object storage with its eleven nines. From a snapshotyou can build a new disk in any zone, copy it to another region for disaster recovery, or make a machine image.A schedule that snapshots every disk nightly and deletes old ones is the baseline backup for virtual machines;AWS Backup, Azure Backup and Google’s Backup and DR service do this across many resources from one place.

After the first full copy, each snapshot stores only changed blocks, so days of history cost little.

File storage

One folder tree, mounted by many machines at once

One shared file system/shared/uploads/invoice-0612.pdfZone aZone bZone call six machines open, read and write the same files at the same timeFile storage is one folder tree that many machines mount at once, across zones.It behaves like the shared drive in an office: paths, folders, permissions, and a file can be locked while someone edits it.It grows and shrinks with what is in it and bills per GB stored. Linux machines reach it with NFS and Windows machineswith SMB. It costs the most per GB, so use it when software needs a shared folder, and object storage when it does not.

A managed file system is a shared network folder with real paths and permissions, open to many machines across zones.

Choosing storage

How the data is reached decides, and the price agrees

Reached throughUsed byYou pay for1,000 GB for a monthObjecta web API, by nameanything, from anywherewhat is stored$23Blockthe operating system, as a diskone machine, same zonethe size provisioned$80Filea network path, as a foldermany machines at oncewhat is stored$300Choose by how the data is reached, then let the price confirm it: object first, block for disks, file last.Good defaults: user uploads, media, backups, logs and analysis data in object storage; operating systems and databaseson block storage; a shared file system only for software that expects one. Example list prices, AWS N. Virginia, 2025:S3 Standard, EBS gp3 and EFS Standard. The ratio between them is similar on Azure and Google Cloud.

Object storage by default, block storage for disks and databases, file storage only when software needs a shared folder.

Data transfer

Moving a mountain of data: sometimes a lorry beats the wire

Sending 100 TB to the cloud over a link used at full speed, day and night100 megabits a second93 days1 gigabit a second9 days10 gigabits a second22 hoursa storage device, shippedabout a week, door to doorAWS Snowball Edge · Azure Data BoxGoogle Transfer AppliancePast a certain size, a box on a lorry is faster than the wire.The provider posts you a rugged, encrypted storage device; you copy the data on, send it back, and they load it intoa bucket. For smaller amounts, or data that keeps changing, online tools copy over the network and resume after breaks:AWS DataSync, AzCopy and Google’s Storage Transfer Service. The bars are computed: 100 TB is 800 trillion bits.

Large migrations use shipped storage devices; smaller or changing data uses online transfer services.

Data transfer · Going deeper

Backups, and keeping storage on site in step with the cloud

Storage in the cloud is durable, and a durable mistake is still a mistake. Backups protect against deletion, corruption and ransomware; hybrid services connect what you still keep on site.

The 3-2-1 rule

3copies of anything that matters
2different kinds of storage
1copy somewhere else: another region or account

One place to schedule backups

  • AWS Backup sets schedules and retention for disks, databases and file systems across an account.
  • Azure Backup does the same from a Recovery Services vault, including on-premises servers.
  • Google Cloud Backup and DR covers virtual machines, databases and file systems.
Replication is not a backup. A replica copies your mistakes within seconds. A backup is a copy from an earlier time that a mistake cannot reach: versioned, and ideally locked against deletion (S3 Object Lock, Azure immutable storage, Google bucket lock).

Hybrid storage

  • AWS Storage Gateway presents cloud storage inside your building as an ordinary file share or disk, caching recent data locally.
  • Azure File Sync keeps Windows file servers on site in step with an Azure file share.
  • Online transfer AWS DataSync, AzCopy and Google’s Storage Transfer Service copy over the network on a schedule and verify what they copied.

Lesson complete

You finished “Storage”

Nicely done.

Open in a new tab