Cloud and platforms
S3 vs Azure Blob Storage vs Google Cloud Storage: Deletes, Versions and Limits
In short
In S3 vs Azure Blob Storage vs Google Cloud Storage, the setting that hurts later is what a delete really does. A plain delete on a versioned S3 bucket only adds a marker, Azure leaves the blob with no current version, and Google keeps deleted objects for 7 days by default even without versioning. All three read strongly after a write.
What does a delete remove in each service?
On a versioning-enabled S3 bucket, a DELETE without a version ID erases nothing. S3 writes a delete marker, that marker becomes the current version, and a later GET answers 404 while the bytes stay stored and billed. Only a DELETE that names a versionId removes data for good, and the console warns that step cannot be undone.
Azure Blob Storage with versioning on behaves close to that, with one visible difference. Deleting without a version ID turns the current version into a previous version and leaves the blob with no current version; there is no marker object. Blob soft delete is a separate switch with its own window of 1 to 365 days, and inside that window Undelete Blob brings the blob back.
Google Cloud Storage stacks two features. Object Versioning keeps a noncurrent copy whenever a live object is replaced or deleted without a generation number. Soft delete is on by default for buckets that support it, holds deleted objects for 7 days, accepts 3 to 90 days, and is switched off by setting the duration to 0.
So a fresh Google bucket already has a week of undo. S3 versioning is disabled until you turn it on, and that is the recovery path its versioning guide describes for accidental deletes.
Checked against: source 2, source 3, source 9, source 10, source 14, source 15
Can versioning be switched off again later?
S3 buckets live in one of three states: unversioned, versioning-enabled or versioning-suspended. Once enabled, a bucket can never go back to unversioned. Suspending stops new versions from piling up, and every version already stored stays where it is.
Azure allows a real off switch. Disabling blob versioning keeps existing blobs, versions and snapshots, and later writes produce a blob that is not a version. You first have to remove any object replication policies on the account. Accounts with a hierarchical namespace for Data Lake Storage are not supported for versioning today.
Google can disable Object Versioning as well. The bucket stops collecting noncurrent versions, and those already present are not touched.
Turning versioning on is not instant on either AWS or Google. AWS treats bucket configuration as eventually consistent and recommends waiting 15 minutes after first enabling versioning before any write or delete. Google asks for at least 30 seconds and warns that deletes may follow the previous setting for a short time. I would build that pause into any script that enables versioning and then starts cleaning a bucket.
Checked against: source 1, source 2, source 9, source 13, source 14
Is consistency the same in S3 vs Azure Blob Storage vs Google Cloud Storage?
For object data the promises line up. S3 gives strong read-after-write for PUT and DELETE in all Regions, and a listing taken after a successful PUT shows the new key. Azure Storage describes a strong model in which reads and lists after an insert or update return the latest change. Google gives strong global consistency for reads after write, metadata update and delete, and for listings, including dual-region and multi-region buckets.
The edges differ. A deleted S3 bucket may still show up in a bucket listing for a while. Google lists access grants and revocations as eventually consistent, and says cached, publicly readable objects might not show strong consistency.
Two writers on one key is where the designs split. S3 has no object locking for concurrent writers: the request with the latest timestamp wins, and AWS says to build locking into your application. Its upload guide does offer conditional writes that refuse to overwrite an existing key. Azure also defaults to last writer wins, but adds ETag checks and blob leases of 15 to 60 seconds or infinite. A write to a leased blob that omits the lease ID fails with 412.
How quickly do lifecycle rules act, and what will they refuse to do?
None of the three fires the moment a rule matches. Adding an S3 rule queues existing objects that already qualify, transitions to the Glacier classes run asynchronously, and billing follows the eligibility date rather than the physical move. Azure says a new or edited policy can take up to 24 hours to take effect and start its first run. Google gives the same 24-hour figure for configuration changes, says the old rules may still act meanwhile, and tells you not to rely on any action happening within a set time.
Each service also has rules that will not run as written:
- S3: objects below 128 KB are not transitioned by default unless a size filter allows it, a single rule cannot move objects out of a class before its minimum storage duration, and nothing leaves Glacier Deep Archive through lifecycle.
- S3 with versioning:
Expirationonly adds a delete marker to the current version;NoncurrentVersionExpirationis what permanently removes the old ones. - Azure: lifecycle cannot rehydrate archived blobs to an online tier, tiering applies only to block blobs, and premium block blob accounts do not support tiering yet.
- Google: a Delete action waits while an object has a hold or has not met the retention policy, and on a versioned bucket deleting a live object only makes it noncurrent.
Checked against: source 3, source 4, source 5, source 11, source 16
How big can one object be, and how hard can one key be written?
Size ceilings are far apart, and so are the per-key rate rules. These figures come from each vendor's current limits pages.
In the table below, the Google rule is the one that catches ported code. A status file or counter object rewritten many times a second will be throttled there, while the S3 and Azure pages give per-prefix and per-blob targets in the thousands. Google buckets themselves start at roughly 1,000 writes and 5,000 reads per second and then scale.
| Limit | Amazon S3 | Azure Blob Storage | Google Cloud Storage |
|---|---|---|---|
| Largest object | 48.8 TiB through multipart upload (10,000 parts of up to 5 GiB) | About 190.7 TiB for a block blob (50,000 blocks of 4,000 MiB) | 5 TiB whatever the write method |
| One request | 5 GB in a single PUT | 5,000 MiB in one Put Blob (service version 2019-12-12 or later) | Multipart uploads of up to 10,000 parts |
| Rate guide | 3,500 writes and 5,500 reads per second per partitioned prefix | Target of up to 3,000 requests per second for one block blob | One write per second to the same object name |
| When pushed too hard | 503 Slow Down while S3 scales | 503 Server Busy or 500 Operation Timeout | Throttling errors above the same-name limit |
What should you check before moving a bucket to another cloud?
I would walk through this list before copying a single object:
- Undo path: S3 versioning with noncurrent expiration maps onto Google soft delete or onto Azure soft delete plus versioning, but the defaults differ, so set retention explicitly on the target.
- Automation that enables versioning: add a wait, 15 minutes on S3 and at least 30 seconds on Google.
- Locking: code that relies on Azure leases needs S3 conditional writes or its own lock elsewhere.
- Hot keys: any object rewritten more than once a second needs a new key layout before it lands on Google.
- Permissions: a revoked Google grant can still work briefly, so do not treat a revoke as an instant cut-off during migration.
Checked against: source 2, source 8, source 13, source 15, source 17
What else do people ask?
Does Google Cloud Storage keep deleted objects when versioning is off?
Yes, if soft delete is enabled, which is the default for buckets that support it. Deleted objects stay restorable for 7 days unless you choose 3 to 90 days, or 0 to turn soft delete off.
Is S3 Object Lock the same as an Azure blob lease?
No. Object Lock keeps objects from being deleted or overwritten for a period, which is a retention control. The S3 consistency notes say S3 does not lock objects for concurrent writers at all.
Sources
- Amazon S3, what is Amazon S3 and the data consistency model, fetched 2026-11-20
- Amazon S3, retaining multiple versions with S3 Versioning, fetched 2026-11-20
- Amazon S3, deleting object versions from a versioning-enabled bucket, fetched 2026-11-20
- Amazon S3, managing the lifecycle of objects, fetched 2026-11-20
- Amazon S3, transitioning objects using S3 Lifecycle, fetched 2026-11-20
- Amazon S3, uploading objects, fetched 2026-11-20
- Amazon S3, optimizing performance, fetched 2026-11-20
- Microsoft Learn, managing concurrency in Blob Storage, fetched 2026-11-20
- Microsoft Learn, blob versioning, fetched 2026-11-20
- Microsoft Learn, soft delete for blobs, fetched 2026-11-20
- Microsoft Learn, Blob Storage lifecycle management overview, fetched 2026-11-20
- Microsoft Learn, Blob Storage scalability and performance targets, fetched 2026-11-20
- Google Cloud, Cloud Storage consistency, fetched 2026-11-20
- Google Cloud, Cloud Storage Object Versioning, fetched 2026-11-20
- Google Cloud, Cloud Storage soft delete, fetched 2026-11-20
- Google Cloud, Object Lifecycle Management, fetched 2026-11-20
- Google Cloud, Cloud Storage quotas and limits, fetched 2026-11-20
How was this page made?
I drafted this page with AI assistance from the official documents listed under Sources. I put this comparison together from the AWS, Microsoft and Google storage documents fetched on 20 Nov 2026; nothing here comes from running buckets myself. Prices, storage-class fees, replication, encryption and SDK behaviour are left out. Found a mistake? Tell me.