Maaz Zindani

Loading the index…

↑↓ move↵ openesc close4133 pages

Cloud and platforms

AWS Secrets Manager vs Azure Key Vault vs Google Secret Manager: Versions, Rotation, Recovery

In short

In AWS Secrets Manager vs Azure Key Vault vs Google Secret Manager, the surprise that costs a weekend is rotation. AWS has managed rotation for some services, while Key Vault and Secret Manager raise an event and leave the new value to your code. Deletes differ as well; AWS waits 7 to 30 days, Key Vault 7 to 90, and a deleted Google secret is gone.

How does each service decide which version is live?

AWS uses staging labels. AWSCURRENT marks the value readers get by default, AWSPREVIOUS the one before it, and AWSPENDING a value that a rotation is still testing. A secret can carry up to 20 labels across its versions, and two versions never share one. Labeled versions are kept; unlabeled ones count as deprecated and are removed once a secret holds more than 100, though not if they are under 24 hours old.

That cleanup rule has a sharp edge. AWS warns that writing a new value more often than every 10 minutes creates versions faster than it removes them, and the quota is 100 versions per secret.

Key Vault has no label layer in the pages I read. Every version gets a 32-character system ID, the newest one becomes current, and a read that leaves out the version returns the latest. Google numbers versions, accepts latest, and lets you add aliases of up to 63 characters, except the names Latest and New.

QuestionAWS Secrets ManagerAzure Key VaultGoogle Secret Manager
What a plain read returnsThe version labeled AWSCURRENTThe latest versionWhatever you name: a number, latest or an alias
Where a half-built value waitsAWSPENDING until rotation moves the labelNo pending marker described; a new version is current at onceA new version that no alias points to yet
Largest value65,536 bytes25 KB64 KiB

Checked against: source 1, source 2, source 9, source 10, source 15, source 16, source 17

Can a reader get an old or expired value?

On Google, yes for a while. Adding a version and then reading it by its number is strongly consistent, but reading through latest or an alias is not covered by that promise. Google says such operations usually settle within minutes and may take a few hours. IAM changes are eventually consistent too, which is worth remembering when a Secret Manager call is denied although the roles look right.

Key Vault treats expiry as advice, not a block. The exp attribute says when data should no longer be retrieved, yet a get still works on expired and not-yet-valid secrets, which the docs suggest for recovery. Code that must refuse an expired secret has to check exp itself.

For a rotation that goes wrong, I would rather pin a Google client to a version number, and on AWS let the previous label carry the old value until the new one is proven.

Checked against: source 9, source 16

Who writes the rotation code in AWS Secrets Manager vs Azure Key Vault vs Google Secret Manager?

AWS gives three routes. Managed rotation, offered for credentials owned by services such as Amazon RDS and Aurora master users, Cognito app client secrets and ECS Service Connect TLS certificates, needs no Lambda function. Partner-held secrets use managed external rotation. Anything else rotates through a Lambda function, and AWS publishes ready-made rotation templates for common cases.

Key Vault's documented pattern is an event pipeline. Thirty days before a secret's expiration date, Key Vault publishes a near-expiry event to Event Grid, which calls a function app you deploy. That function makes the new password, stores it as a new version and updates the target system. If a step fails, Event Grid retries for two hours.

Google only schedules. You set a rotation_period of at least one hour and a next_rotation_time at least five minutes ahead, and Secret Manager sends a SECRET_ROTATE message to the Pub/Sub topics on the secret. A subscriber you build must create the new version. Google's soft limit for adding versions to one global secret is 2 requests per second and 120 per minute.

Checked against: source 2, source 3, source 4, source 5, source 11, source 14, source 17

What happens after someone deletes a secret?

Here the three are furthest apart. Read each line as what you still have the morning after a mistaken delete:

  • AWS: the secret turns inaccessible at once and is scheduled for deletion after a recovery window of 7 to 30 days, 30 by default. restore-secret works until the DeletionDate. --force-delete-without-recovery skips the window. A primary that is replicated cannot be deleted until its replicas are removed, and a deleted replica goes immediately.
  • Azure: soft delete is on by default for new vaults and cannot be turned off once enabled. Deleted secrets stay recoverable for 7 to 90 days, 90 by default, and that interval is chosen at vault creation and cannot change later. Purge protection is optional and off by default; when on, nothing is purged before the interval ends.
  • Google: deleting a secret is irreversible. Destroying one version is permanent too, unless delayed destruction is enabled, in which case the version is disabled and scheduled, and an admin can restore it meanwhile. Disabling a version is the step you can undo.
  • So on Google I would disable first and delete later, as a habit rather than a feature.

Checked against: source 6, source 7, source 12, source 19, source 20, source 21

Where are copies kept, and what happens in a regional outage?

AWS replication is opt-in per secret. You pick Regions; each replica has the same ARN apart from the Region, a new value propagates to every replica, and a replica can be promoted to a standalone secret if needed.

Key Vault replicates for you. Data is copied synchronously across availability zones where the region has them, and asynchronously to the paired region when that pair sits in the same geography. A regional failover is started by Microsoft, on a best-effort basis and possibly hours after the outage begins. After it, the vault is read-only: no property, access policy or firewall changes. Some regions have no cross-region replication at all.

Google asks once, at creation. Automatic replication spreads the payload across multiple regions globally; user-managed replication keeps it in the locations you list, which matters when an organisation policy limits where data may live.

Checked against: source 8, source 13, source 18

What should you settle before picking one?

Four questions decide most of it:

  • Is the credential owned by a service with AWS managed rotation? If yes, AWS removes the rotation code entirely; on the other two you write and run it.
  • Do you need writes during a regional outage? A failed-over Key Vault is read-only, so plan a second vault if rotation must keep running.
  • How long should a deleted secret survive? Key Vault fixes the window at vault creation; Google has no window for a whole secret.
  • Do readers use latest? On Google that path is eventually consistent, so pin numbers during a cutover.

Checked against: source 4, source 11, source 12, source 14

What else do people ask?

Can Key Vault soft delete be disabled on an existing vault?

No. Once soft delete is on, Microsoft says it cannot be turned off, and the retention interval set at creation cannot be edited later.

Does Google Secret Manager rotate a database password by itself?

Its rotation feature sends a SECRET_ROTATE message to Pub/Sub on your schedule. Changing the password and adding the new version is the job of a subscriber you run.

Sources

How was this page made?

I drafted this page with AI assistance from the official documents listed under Sources. Written from AWS, Microsoft Learn and Google Cloud pages read on 20 Nov 2026. I did not operate these three secret stores to write it, so every behaviour here is what the vendors document. Pricing, keys and certificates, HSM tiers and per-language SDKs are out of scope. Found a mistake? Tell me.

Keep reading

Next step

Have a project in mind?

Tell me about it. A discovery call within 48 hours.