A Long Peak-Season Freeze Moves Risk into January: Check Your Release Safety in 30 Minutes

A note to forward to your team

Peak season release check

Black Friday is 27 November and Cyber Monday is 30 November. Before we agree a freeze window, I would like a few numbers.

Please run the checks in this note by [date] and reply with:
1. Changes merged during last year's freeze
2. Changes in the first release in January
3. Measured rollback time of [revenue service], in minutes
4. Irreversible schema changes planned before January

Owner: [name]
Note: https://stonetusker.com/insights/peak-season-release-readiness.html

Why it matters now

The busiest days of the year are close. Adobe expects the five days from Thanksgiving to Cyber Monday to bring 47.5 billion dollars of online sales, 17.3% of the season. Cyber Monday is forecast at 15.1 billion dollars and Black Friday at 12.9 billion dollars.² Travel, gaming and streaming businesses have their own peaks, such as fare sales, game launches and live events, and the same reasoning applies.

A freeze from mid-November into January feels safe. It stops deployments and leaves development running, so finished work piles up and the first release after the freeze carries weeks of change at once. Research from the DevOps Research and Assessment programme shows that speed and stability go together.³ Small batches reach feedback sooner, and when one fails, the cause is easier to find.⁴

The goal is a release that is small, observable and quick to reverse. With that in place, a short freeze around the busiest days is a business choice backed by data, and the weeks around it stay productive.

Check your release safety in 30 minutes

Step 1. Look at the rollout history

Start with the service that carries revenue at peak, such as checkout, booking or playback. On Kubernetes:

kubectl rollout history deployment/checkout --namespace shop

Step 2. Time a rollback

In a rehearsal environment, time a rollback from the moment you decide to the moment traffic is healthy again.

kubectl rollout undo deployment/checkout --namespace shop

Step 3. Count last year's freeze backlog

Inside a clone of the main repository, count the changes merged during last year's freeze.

git log --merges --oneline --since=2025-11-15 --until=2026-01-10 | wc -l

If your team squashes pull requests into single commits, run the same command without --merges.

Step 4. Answer four questions in writing

• How long does a rollback of the revenue service take, measured rather than estimated?
• Which share of last month's changes reached customers behind a feature flag?
• Which schema changes planned before January are irreversible?
• Is there an error budget for the revenue path, and does it decide when releases stop?

What the result means

An error budget turns the decision to stop into a rule. In the example policy from the Google site reliability engineering workbook, a service that exceeded its error budget for the preceding four-week window halts changes other than the most urgent (P0) issues and security fixes until it is back within its objective.⁹

The preparation, step by step

1. Measure the rollback. Time a rollback of each revenue service in a rehearsal environment, and fix the slowest step first.
2. Put risky changes behind flags. Wrap every change planned for November and December that alters customer-facing behaviour in a feature flag with a named owner.
3. Finish irreversible migrations early. Complete or postpone any irreversible schema change, and split the rest into expand, migrate and contract phases.
4. Automate canary analysis. Route a small share of traffic to each new version and abort automatically when error rate or latency moves beyond agreed limits.
5. Agree an error budget policy. Write down when releases stop and restart based on reliability data, and agree a short freeze only for the busiest days.
6. Rehearse an incident. Run a game day on the revenue path with the on-call team before Thanksgiving, including a rollback under load.

The tools are common. Argo Rollouts aborts a rollout and returns traffic to the stable version when analysis fails, and Flagger rolls back when failed checks reach a threshold.^5,6 Feature flags separate releasing a feature from deploying the code.⁷ Expand and contract splits a schema change into three phases, so the application can roll back at every step.⁸

What is still open

• The spending figures are forecasts. Adobe measures United States online sales for November and December, and other forecasters use different scopes and periods.
• A short freeze around the busiest days can still be the right decision, for example for staffing or contractual reasons.
• These practices depend on reliable service level indicators, so teams that lack them should start there.