One API key may serve a website, an automation connection and a scheduled job. Revoking it before finding every consumer can break workflows that appeared unrelated. The decision framework below works for small teams in the US, Canada or UK, but the vendor’s actual key limits, logging and recovery rules decide the execution plan.
The short answer
Inventory consumers before choosing a rotation sequence. If two valid credentials can overlap, create the new key, migrate consumers, verify usage and only then revoke the old one. If the provider allows one key only, plan a change window and recovery route; do not promise zero downtime. If the old key is exposed, containment may require immediate revocation even if availability suffers.
Google Cloud notes that a rotation notification does not itself replace a credential in an external system. Microsoft Entra recommends adding the new credential, updating the service, validating which key ID is used and only then removing the old one. Paddle documents a parallel-key path for planned rotation and a different priority for exposure. These are vendor examples, not a universal guarantee.
Pick the right regime
| Confirmed situation | Decision | Stop condition |
|---|---|---|
| Two keys can coexist; no suspected leak | Use a brief overlap, migrate consumers and verify traffic on the new key | Do not revoke while a consumer is unknown or still uses the old key |
| One key only, or rotation invalidates the old key immediately | Schedule a window, owner, test and recovery path; communicate possible interruption | Do not claim “no downtime” without a supported transition |
| Key may be compromised | Treat it as an incident; contain or revoke under the provider’s guidance and investigate usage | Do not keep a compromised key active just to preserve overlap |
This is Pulse’s decision tree. A routine availability-first rotation sequence is not the right response to a leaked credential.
Inventory without storing the secret
Record the credential’s name or ID, provider, environment (test or production), owner, permissions, expiry, consuming systems, affected jobs, change owner and way to verify usage. Never put the key’s value in the inventory, a shared spreadsheet, ticket, URL or logs. Include infrequent jobs: a monthly export will not prove migration during a five-minute check.
| Consumer | Where to verify the update | Evidence of use | Failure response |
|---|---|---|---|
| Website or app | Secure configuration and deployed version | Authorized request attributed to the new key | Revert deployment if the old key is still valid |
| Automation platform | Connection and every workflow reusing it | Legitimate runs without authentication errors | Pause dependent action and repair connection |
| Scheduled job | Secret store and next run | The next scheduled run completes | Follow the contingency procedure |
Use an authorized controlled test or observe a legitimate run; do not manufacture business activity just to make a dashboard move. The observation window must cover the least frequent consumer. Check that the new key has only the permissions it needs.
Planned rotation when two keys can overlap
- Confirm in the provider’s current documentation that both credentials can coexist and how revocation works. Assign an owner and observation window.
- Create the new key in the correct environment. Store it in the approved secret-management mechanism, never in code or logs.
- Update one consumer at a time, recording version, time and outcome. A process may cache a credential until restarted.
- Check successful operations and, when supported, the key ID in authentication logs. Watch both routine and scheduled work. No error is not proof that the new key is being used.
- Revoke the old key only after the inventory is complete and the planned window shows migration. Monitor authentication failures afterward with a named responder.
If the new key fails before revocation, pause the migration and restore the previous configuration only if the old key remains valid and uncompromised. After revocation, recovery may require another new key. Paddle has provider-specific reactivation rules; do not assume they apply elsewhere.
Six checks before retirement
Review hypothetical cases: a successful daily task, a monthly job not yet run, a connection shared by several workflows, a test key mistakenly placed in production, a 401 after the change, and a possibly leaked key. Do not revoke while the monthly job lacks a validation plan, the shared connection is unmapped, the environment is wrong or no one owns an authentication failure. For exposure, reverse the priority: contain the risk and follow incident response even if work pauses. These are proposed checks, not outcomes observed by Pulse.
The periodic access review helps identify who should retain access. The automation failure contingency plan covers recovery if a workflow breaks. Reviewing people does not replace mapping technical consumers of a credential.
Sources and limits
Google Cloud’s rotation guidance separates a reminder from actually updating the application. Paddle’s API key guide describes overlap and the exposure exception. Microsoft Entra recommends validating the new credential before removing the old one. The decision tree and inventory are our synthesis, not a zero-downtime guarantee or a field-tested deployment. Check your provider and organizational policy before changing production.