Here’s a Saturday we made up. Some of you will recognize it. The freeze started Wednesday night. Deploys are locked until the Monday after Thanksgiving, the on-call rotation is posted, and the release manager’s email has been skimmed by everyone it went to. A product manager checking numbers over coffee sees the new recommendations panel still sitting at 10%. The conversion figures look good. They move it to 100% from their phone. The panel makes three extra calls per page view, to a service sized for a tenth of peak traffic.
No deploy happened, so the CI lock never fired. Separating shipping code from releasing it is why teams adopt flags in the first place, and it also means a freeze that only locks CI leaves the release switch within reach of anyone with dashboard access.
Key Takeaways
- A flag change releases code without a deploy, so a freeze that only locks the pipeline misses it.
- In production, freeze the changes that move users between variations: toggles, targeting rules, rollout percentages and copied configuration.
- Let automated kill switches fire. People, including someone turning a flag off, go through a break-glass step with a written reason.
- Before the freeze starts, approve the launches that have to land inside it and settle every flag that is halfway rolled out.
1. Why flags slip past a code freeze
Teams freeze when an outage would cost the most and the fewest people are around to fix one. For US retail that means Thanksgiving through Cyber Monday, and some teams stretch it to New Year.
Most of these freezes live in CI. The deploy job gets disabled, merges to main need a second sign-off, or a bot asks on every pull request whether this really can’t wait until December. For years that was enough, since a deploy was how production changed.
With flags, the recommendations panel’s code went out weeks ago, dark. What users get depends on a value in the flag platform that anyone with access can change in a few seconds from a browser tab. That’s how dark launches and instant rollback work, and during a freeze it means releases keep happening where the CI lock can’t see them.
A scheduled change somebody set up in October can fire on the busiest morning of the year. Or a teammate copies staging’s targeting into production to fix one small rule, two unfinished rules ride along with it, and nobody notices until Monday’s numbers come in.
2. Decide what the freeze covers
Ask a team to “leave the flags alone” and the first question back is whether that includes turning one off. Write the list down instead. Sort every kind of change by a single test, whether it moves any user from one variation to another, and the table is how that comes out for the common cases.
| Change | Freeze it? | Why |
|---|---|---|
| Turning a flag on or off | Yes | Changes what every user in the environment gets |
| Editing, adding or reordering targeting rules | Yes | Moves users between variations |
| Raising or lowering a rollout percentage | Yes | Moves users between variations |
| Copying configuration from staging into production | Yes | Several of the changes above at once |
| Advancing an automated rollout | Yes | The same as raising the percentage, done by a timer |
| Automated kill switch firing | No | Needs to work most when things go wrong |
| Reading flags and SDK evaluation | No | Your application depends on it |
| Renaming a flag, editing its description, setting an owner | Usually no | Changes no user’s experience |
Segments need a check of their own. In Featureflip a segment is shared across environments, so editing one can change who matches a production rule while production is frozen, and the freeze won’t stop it. If your platform works the same way, put segment edits in the freeze announcement as a team rule.
Freeze production and leave staging open. The week before Black Friday is when people want to try the January launch somewhere safe, and blocking staging just pushes that work into December. If every project names its environment production, one window matched on that key covers all of them.
3. Keep the kill switch working
Picture the checkout throwing errors at 11:00 on Black Friday, with the flag that would stop it frozen like everything else.
Could you just allow every turn-off, since off sounds like the safe direction? Take a flag that has been at 100% for two months and now guards the only checkout path that works with this year’s payment provider. Turning it off sends every shopper to the old path, which nobody has run in production since September.
Split it by who is making the change instead. A flag trigger wired to an error-rate alert was set up on a calm day, for one flag and one failure, and that kill switch should fire straight through a freeze. (We wrote up which alerts deserve one.) A person changing a flag, in either direction, goes through break-glass and writes down why. During a real incident that costs about ten seconds, maybe a few minutes if the person with the role is on a plane. On the made-up Saturday, the product manager hits a dialog telling them to find an admin, and the panel stays at 10%.
4. Launches that belong inside the window
The Black Friday banner goes live at midnight on Friday, and the holiday pricing page switches on at 06:00 Monday. Both fall inside the freeze on purpose. Scheduled changes take care of the timing, which leaves the question of what the freeze does with a change when it comes due.
The safer default holds the change until the freeze ends and applies it then, which suits the routine edit someone scheduled back in October without checking a calendar. Some changes do harm if they land late, though. A discount meant to end with the sale is the usual example, and for those a failure that names the freeze is better, so the owner knows to reschedule.
Approve the launches themselves before the freeze, one at a time, by someone who could break the freeze anyway. Tie each approval to the window it was given for. If an incident on Friday afternoon opens a second freeze on top of the first, the midnight banner should wait until somebody looks at it again.
5. Break-glass without a back door
Leave break-glass out of the plan and an admin will improvise it by ending the freeze early and opening a new window after their change. All the audit trail shows is a window that closed early with no reason given.
Keep the exception to owners and admins (or whoever your escalation policy names), and tell a developer who hits the freeze who to ask. They type a reason at the moment they break it. “INC-412, reverting the 14:02 change” tells a reviewer in January what happened, and a reason that just says “fix” is worth raising at the retro. The override covers one change and the freeze closes again behind it, so nobody has to remember to re-arm anything. The audit log records who broke which window and why, and a webhook to the incident channel lets the rest of the team see it that day.
If your platform already asks for a change reason in production, make sure a routine reason can’t break a freeze by accident. Featureflip’s API takes the reason in an X-Intent-Reason header and wants a second header, X-Freeze-Override: true, before it lets the change through.
curl -X POST "https://api.featureflip.io/api/v1/orgs/acme/projects/checkout/flags/new-checkout/environments/production/toggle" \ -H "Authorization: Bearer $FEATUREFLIP_TOKEN" \ -H "Content-Type: application/json" \ -H "X-Intent-Reason: Reverting the 14:02 deploy, INC-412" \ -H "X-Freeze-Override: true" \ -d '{"enabled": false}'Leave the override header off and the request gets a 423 with a change_frozen error naming the freeze and when it ends, which a script or an agent can pass on to whoever ran it.
6. Rollouts that are halfway through
Any flag sitting between 0% and 100% in production on Wednesday night will spend five days of peak traffic at that number unless someone decides otherwise. Go through them in the planning meeting, with each flag’s owner there.
A rollout that has looked healthy for a week can go to 100% now, and holiday traffic becomes its last test. Pull anything that touches checkout and hasn’t been watched closely back to 0%, and bring it back in December. An automated rollout can also pause when its next step falls due inside the freeze and carry on afterwards, as long as you’re comfortable running at its current percentage through the peak.
7. A freeze checklist for the holidays
We’d spend the week before the freeze on this, roughly in this order.
- A week out: list every flag mid-rollout in production, plus every scheduled flag change due between the freeze start and two days after it ends. That list is the agenda for one short meeting.
- Five days out: decide each mid-rollout flag. Approve the scheduled changes that are real launches and leave the rest to the hold policy.
- Three days out: create the freeze window on production, with a name and reason people will see when they’re blocked (“Black Friday freeze, ask #oncall for an override”), and confirm the triggers on the riskiest flags still point at live alerts.
- The day before: send the announcement with the window times and who can break the freeze. Spell out the team rules for anything the freeze can’t block, segment edits included.
- During: let the alerts and the approved launches run.
- After: read every override and pass-through in the audit log at the retro. A launch nobody remembered to approve or a flag someone left at 40% goes on next year’s list.
8. Change freezes in Featureflip
Featureflip has change freezes built in on the Business and Enterprise plans. You create a window in Organization settings, scheduled ahead or with Freeze now for an incident, and pick the environments it covers. An organization-wide window matched on the production key covers that environment in every project.
While a window is active, any attempt to change what a frozen environment serves is refused, whether it comes from the dashboard, the Management API, the CLI or the MCP server, and SDK evaluation and streaming carry on as normal. Flag triggers pass through and are logged as pass-throughs. Rollout ramps pause when a step falls due and pick up again once the window ends, and scheduled changes follow the window’s Hold or Fail policy unless an Owner or Admin has pre-approved one through that specific window. Windows also send change_freeze.started and change_freeze.ended webhooks, so a Slack channel can announce them.
Freezes sit next to rollout ramps, flag triggers and required change reasons on the release safety page. For how an auditor reads the same controls, see feature flag governance.
The shorter version
If you do one thing this week, open the list of flags sitting between 0% and 100% in production and decide each of them before Wednesday. Then create the window on production, approve the launches that belong in it, and leave the kill switches armed.
Frequently asked questions
Should feature flags be frozen during a code freeze?
Yes, in production. A flag change alters what users get without a deploy, so a freeze enforced only in CI leaves every flag open. The test for what to freeze is whether a change moves any user from one variation to another. Reads, SDK evaluation and automated kill switches keep running, and staging can usually stay open.
Can you turn off a feature flag during a code freeze?
Through a break-glass step, yes. Turning a flag off can be risky when an old, fully rolled-out flag guards the only code path that still works, so a person should give a reason and use an explicit override, which the audit log records. An automated kill switch wired to an alert can fire on its own.
What happens to scheduled flag changes during a freeze?
It depends on the freeze’s policy. Under a hold policy the change waits and applies once the freeze ends, while a fail policy marks it failed with a reason naming the freeze. In Featureflip an Owner or Admin can also pre-approve a single scheduled change through one specific window, for a launch that has to go live at a set time.
What is the difference between a code freeze and a change freeze?
A code freeze usually stops deploys. Change freeze is the older IT operations term for holding any change to production, configuration and feature flags included. On a team that ships behind flags most user-facing changes skip the deploy step, which makes the wider definition the useful one.
Does a change freeze stop my app from getting flag values?
No. A freeze blocks writes to flag configuration in the environments it covers. Reading flags, evaluating them and SDK streaming continue as normal, so applications keep receiving values through the whole window.
Featureflip is a flat-priced feature flag platform with change freezes, rollout ramps, flag triggers, required change reasons, and a cleanup Action that opens a pull request to remove each dead flag on GitHub, GitLab or Bitbucket. Start on the Free plan without a credit card. Change freezes come with the Business plan.