Feature Flags Explained: Release Changes Gradually
What are feature flags used for?
Feature flags are controls that let an application choose whether someone receives a particular feature or behaviour. They let a team deploy the code for a change separately from making that change available to customers. They are used for gradual releases, early access, controlled comparisons and emergency switches.
For a founder, the useful question is simple: can we introduce this change to a limited audience, see what happens, and stop further exposure if something goes wrong?
How does a feature flag work?
Imagine replacing a product's existing search screen. With a flag, the application checks a rule before deciding which screen to show.
When the flag is off for a customer, they see the existing search. When it is on, they see the replacement. Both paths can be present in the deployed application, while the rule determines access.
That creates separate decisions:
- Deployment: Is the code ready to run?
- Exposure: Who should receive the change?
- Expansion: What evidence would justify offering it more widely?
Ask your build partner to explain each decision separately. “The code is deployed” should not leave you guessing whether customers can already use it.
How should you release a change gradually?
Start with a clearly defined group and a reason for choosing it. For a shared workspace, consider giving the whole account the same experience so colleagues can follow the same instructions.
The following rollout is illustrative, not a required sequence.
| Stage | Who receives the change | What to check before expanding |
|---|---|---|
| Product review | People checking the completed feature | The main task works and the wording is clear |
| Early access | Customers who agreed to try it | Real tasks succeed without unexpected obstacles |
| Wider access | More eligible accounts | Support reports and technical results remain acceptable |
| General availability | Everyone intended to receive it | Remaining limits are understood and documented |
Define the evidence before starting. For a search replacement, that might mean customers can find the expected records, filters return the right results, and saved links still open correctly.
Keep a customer's assignment stable during the review where possible. Switching their experience repeatedly makes feedback harder to interpret. Decide whether assignment belongs to a person, an account or a workspace, based on how customers use the product.
What can an emergency switch actually stop?
An emergency flag is intended to stop new use of a problematic path without waiting for another code deployment. Its effectiveness depends on where the check happens and how quickly the application receives the changed setting.
Before relying on one, ask:
- Which actions will stop when the switch changes?
- What happens to requests already underway?
- Will background work continue?
- What will customers see instead?
- How will we confirm the switch has taken effect?
“Turn it off” needs a defined outcome. A useful fallback might restore the previous screen or temporarily prevent a particular action with a clear explanation.
For a critical operation, request a demonstration of the switch and its fallback before launch.
Why does disabling a flag not undo every change?
A flag controls which behaviour runs. It does not automatically reverse everything that behaviour has already done.
If a feature sends a message, switching it off does not recall that message. If it writes records, changes account settings or triggers another service, those effects need their own recovery plan.
Consider an illustrative address-cleanup feature. It rewrites saved delivery addresses into a new format. Turning it off might prevent further rewriting, but the addresses already changed remain changed.
Before enabling such a feature, establish:
- Whether the original information is retained.
- Whether the previous behaviour can read the changed records.
- How incorrect changes would be identified and repaired.
- Whether affected customers need an explanation.
A switch is most useful when the surrounding design makes stopping safe. Review data compatibility and recovery alongside the flag.
Who should have permission to change flags?
Treat permission to change a flag as permission to change the customer experience.
Recommend a named owner for each flag, limited access to important controls, and a record of changes. For switches affecting sensitive actions, decide who can authorise a change and who can act during an urgent incident.
Keep two kinds of permission separate: permission to operate the flag, and permission for a customer to perform the action.
Hiding a button with a flag is not sufficient access control. The application must still check whether the customer is authorised when they request the action.
This distinction matters for paid features, private records and account administration. The broader principles are covered in data security in modern applications.
What needs testing before a flag goes live?
Request evidence for both enabled and disabled behaviour. The existing path remains part of the product while customers can still receive it.
Use this acceptance checklist:
- Enabled: The intended audience can complete the new task.
- Disabled: Customers can still complete the existing task.
- Eligibility: Customers outside the selected group receive the expected experience.
- Authorisation: Direct requests cannot bypass access checks.
- Failure: An unavailable flag setting produces a deliberate, tested default.
- Transition: Changing the setting during an active task has a defined outcome.
- Data: Both paths handle the records they may encounter.
Choose the default for each feature deliberately. Restoring an established search screen may be reasonable; enabling a sensitive action when its eligibility cannot be checked may be inappropriate.
For more on defining useful acceptance checks, see software testing and quality assurance.
What common mistakes should founders watch for?
Expanding without a decision rule. Agree what success looks like and what would stop the release. Otherwise, gradual exposure becomes a sequence of guesses.
Assuming “off” means recovered. Separate preventing further effects from repairing existing ones.
Measuring only technical errors. Include the customer's task. A screen can load successfully while making the intended action confusing or impossible.
Combining unrelated changes behind one switch. Prefer a flag with a clear purpose and a predictable fallback.
Keeping every flag forever. Temporary flags leave alternative paths to understand and maintain. Schedule their removal as part of finishing the feature.
Adding a flag without a useful choice. If there is no meaningful audience restriction, fallback or stopping action, ask what the extra control achieves.
When should a feature flag be removed?
Give temporary flags an exit condition when they are introduced. That might be successful general availability, rejection of the replacement, or completion of a limited trial.
Once the decision is settled, recommend removing the unused behaviour, retiring the flag setting and updating the relevant tests. Confirm that nothing still depends on the setting before deleting it.
Some controls may be deliberately permanent, such as an operational switch. Keep those documented and test them periodically. A permanent control should have an explicit purpose.
Unfinished cleanup belongs in the product's maintenance decisions; technical debt explained for founders provides useful context.
What else do founders ask about feature flags?
Are feature flags the same as experiments?
No. A flag controls exposure. An experiment also needs a defined question, a comparison method and a way to assess the result. Giving different groups different experiences does not, by itself, establish which experience is better.
Should every feature have a flag?
Use one when controlling exposure solves a specific problem. Ask what decision the flag enables, what happens when it is off, and when it can be removed.
Can a flag replace a recovery plan?
No. It may stop further use of a path, but completed actions and changed records can require separate repair.
What should I ask a software partner to show me?
Ask for a demonstration of both states, the audience rule, the fallback, the permissions and the recovery limits. Request a clear owner and an exit condition before treating the feature as ready.
*Planning a change that needs a careful introduction? Start a conversation about the audience, fallback and acceptance checks.*