Quick answer
A code freeze is a period when changes to a system's code are stopped or tightly restricted, usually before a major release or a seasonal peak. It does not remove risk, it concentrates it in the release that opens and the one that closes the freeze, so fix your baseline numbers before it starts.
Formula: Error-rate ratio = error rate in the window after a release ÷ baseline error rate (previous 7 days); flag when the ratio is at least 2 and the window has at least 30 errors
What is a code freeze?
A code freeze is a period in which changes to the code of a system, or to part of it, are stopped or tightly restricted. Teams usually start one shortly before a major release or a seasonal traffic peak, and end it when the release is stable or the peak is over. The aim is to stop new changes from landing at the moment a failure would cost the most.
A freeze does not remove risk. It moves it. Changes that would have shipped one by one pile up behind the freeze and then ship together, and the change you are most afraid of is the one made on the last day before the freeze starts. The risk gets concentrated into two moments: the release that opens the freeze and the one that closes it.
That is why a freeze is a measurement problem as much as a process one. If you know what normal looks like before the freeze starts, you can tell within an hour whether a release broke something. If you don't, every wobble in the numbers turns into an argument about whether it was the release, the weather or the traffic. The rest of this page covers the types of freeze, the five numbers worth watching around one, a worked example you can recompute, and the mistakes that turn a freeze into theatre.
Types of freeze
The word covers several different practices. Agree on which one you mean before you announce it:
| Type | What is blocked | What is still allowed | Typical length | Main risk |
|---|---|---|---|---|
| Full code freeze | Every change to the system's code | Nothing, except an approved emergency fix | Hours to a few days | Pressure to rush work in before it starts |
| Soft freeze (feature freeze) | New features and refactors | Bug fixes and low-risk changes after review | Days to weeks | "Bug fix" quietly becomes a new feature |
| Holiday or seasonal freeze | Non-essential releases during a peak such as a sales season | Security patches, critical fixes, agreed content changes | One to several weeks | Large backlog released all at once afterwards |
| Specification freeze | Changes to requirements | Implementation of what was agreed | Until the build phase ends | Late discovery that the spec was wrong |
A feature freeze and a specification freeze are older terms from release planning; a code freeze is the strictest of the three. When people say "freeze" in conversation they often mean the soft version, so write down, in one sentence, what is allowed. "Only fixes for the P1 list, reviewed by two people" is a rule. "Be careful" is not.
Why teams freeze, and the case against
The case for a freeze is simple. Fewer changes mean fewer new failure causes, and the team's attention goes to watching the system instead of changing it. For a retailer, a payment service or a ticketing site, the cost of a failure is much higher during a peak than on an ordinary Tuesday.
The case against is also reasonable. The UK Ministry of Justice's technical guidance tells its teams to avoid code freezes where possible. It lists five costs: exception processes that nobody practises, rushed releases before the freeze, work that is stale by the time it ships, large batches released together after the freeze and vague scope that drifts. Its alternative is smaller, more frequent releases that make any one change less frightening.
For the cases where a freeze still makes sense, the same guidance gives six conditions: it covers one system or a small group maintained by one team, it lasts at most a few days, no other work is planned during the window, confidence in releasing is low and cheaper to work around than to fix, follow-up work is planned to remove the need for the next freeze, and the exception process is documented. If you cannot meet all six, the guidance's advice is to make the system safer to change instead.
You do not have to agree with all of that to take one lesson from it. A freeze is a bet that the risk of changing is higher than the risk of not changing. Numbers tell you whether the bet is paying off.
The five numbers to watch around a freeze
Watch the same five numbers before the freeze, during it and after it. Define each one against a baseline you set before the freeze starts. The thresholds below are sample starting points, not standards. Tune them to your own traffic.
| # | Number | Baseline | Flag it when | Why this one |
|---|---|---|---|---|
| 1 | Error rate (errors ÷ pageviews) | Mean of the previous 7 days | At least 2 times the baseline and at least 30 errors in the window | The most direct sign that a release broke something |
| 2 | Pageviews | Same weekday, recent weeks | More than 30% above or below | A drop can mean a broken page or a broken tracking snippet; a spike means your capacity is being tested |
| 3 | Goal conversion rate | Mean of the previous 14 days | Below the baseline minus two standard errors | Catches breakage that raises no error, such as a button that no longer submits |
| 4 | Largest contentful paint, 75th percentile | Mean of the previous 7 days | Above 2.5 seconds, or 20% above your own baseline | A release can slow pages without throwing an error |
| 5 | Minutes with no data | Zero | More than 15 minutes | A release can take the tracking snippet down, and silence looks like a traffic drop |
Three details keep these rules honest:
- Set the baseline before the freeze. A baseline taken after a release includes the problem.
- Require a minimum volume. A 2-times ratio on four pageviews is nothing. The 30-error floor in row 1 is there for this reason.
- Mark what you changed and when. Without dates on the chart you cannot separate your release from a campaign or a holiday.
The 2.5-second threshold is the "good" boundary Google's web.dev documentation gives for largest contentful paint, measured at the 75th percentile of page loads. Everything else in the table is your own policy.
Worked example: when does 1.6% mean "act"?
Sample data: a store's 7-day baseline before the freeze is 70,000 pageviews and 560 errors, so the baseline error rate is 560 ÷ 70,000 = 0.8%. A release goes out at 14:00 on the day the freeze opens.
One hour later, the one-hour window shows 5,000 pageviews and 80 errors, an error rate of 80 ÷ 5,000 = 1.6%. The ratio to the baseline is 1.6 ÷ 0.8 = 2.0.
How do we know that is not chance? At a baseline of 0.8%, 5,000 pageviews should produce about 40 errors. A Poisson model, which fits rare events like this, gives the chance of seeing 80 or more by luck as roughly 2 in 100 million. That is not noise. The rule from the table fires: ratio at least 2, errors at least 30.
Now compare three windows with the same release:
| Window | Pageviews | Errors | Rate | Ratio to 0.8% | Expected errors | Chance of this by luck | Decision |
|---|---|---|---|---|---|---|---|
| A | 5,000 | 80 | 1.6% | 2.0 | 40 | about 0.000002% | Act: roll back or hot-fix as agreed |
| B | 5,000 | 48 | 0.96% | 1.2 | 40 | about 12% | Watch, check again in an hour |
| C | 250 | 4 | 1.6% | 2.0 | 2 | about 14% | Not enough volume; wait for 30 errors or more traffic |
Window C has the same 1.6% as window A and means almost nothing, because four errors can easily happen by luck when you expected two. This is why the rule has two parts, a ratio and a volume floor. One caveat: errors tend to come in bursts, for example one visitor hitting the same bug ten times, so real variation is wider than the Poisson model says. Treat the probabilities as a way of seeing why volume matters, not as exact odds.
The conversion row in the table works the same way. With a 3.0% baseline and 2,000 visitors in the window, you expect 60 conversions and the standard deviation is about 7.6. Two standard errors below is about 44.7, which is 2.24%. Fewer than 45 conversions out of 2,000 is worth a look.
What to decide before the freeze starts
A freeze needs five decisions in writing before day one. Every item on this list is something teams argue about at 2 a.m. if they have not settled it:
- Scope. Which systems and which kinds of change, including configuration, feature flags, content and third-party scripts. A freeze on code that leaves marketing free to push a new tag manager container is not a freeze.
- The emergency lane. Who can approve a fix, what counts as an emergency, and how many reviewers it takes. Keep it short and practise it once.
- The baseline. The five numbers above, with their baseline values and flag rules written down, so the person on call does not have to decide on the spot.
- The thaw criteria. What has to be true for the freeze to end: for example, a stable error rate for 48 hours after the last release, or a calendar date. "When it feels safe" is not a criterion.
- The release plan for after. Whether the backlog goes out as one batch or several smaller ones. Smaller is easier to diagnose.
Related practices cover the same ground from different angles. Canary testing limits the damage of a single release by sending it to a few users first; this page is about what to watch while the whole system is frozen. A good error rate definition is the basis of the first row, and anomaly detection is the general method for flagging a number that has left its normal range. For page speed in row 4, see web performance monitoring.
Common mistakes
- Sneaking in hot fixes. Each "tiny" change made during a freeze skips the process the freeze exists to protect. Use the emergency lane, log it and mark the date on your charts.
- No thaw criteria. A freeze with no end condition drifts, spreads to neighbouring systems and gets extended. Write the end rule on day one.
- Taking the baseline after the first release. Then the problem is in the baseline and the ratio looks fine.
- Watching totals only. A broken checkout button can leave pageviews and errors unchanged. Conversion and the no-data check catch what the error rate misses.
- Treating the freeze as proof of safety. A quiet freeze says the system survived with no changes. It says nothing about the release that ends it.
- One giant release at the end. Held-back changes interact with each other, which makes any failure harder to trace. Release in smaller groups and watch each one.
- Forgetting the people who are not developers. Content edits, tag changes, pricing updates and campaign launches can break pages as surely as code. Put them in scope or agree what is exempt.
Watching a freeze in TapCub
Take the sample store above and set it up in TapCub the week before the freeze.
- Mark the dates. In Configuration › Annotations, add a date-range annotation titled "Code freeze" (type Other) and a single-day Release annotation for each release. You can also bulk-import up to 500 rows from a CSV with
date_from,date_to,type,title,noteandscopecolumns. Annotations appear on the trend chart in Web analytics › Overview, so a jump has a reason next to it. Creating them needs the analyst role or higher. - Set the baseline. Open Monitoring › Errors with the previous seven days selected and note the error count, the error rate (errors ÷ pageviews) and the number of affected sessions. Error collection has to be turned on in the install snippet, otherwise an empty list does not mean your site threw nothing. Do the same for the 75th percentile of largest contentful paint in Monitoring › Performance.
- Create monitors. In All-platform analytics › Operations › Monitors, create a fixed-threshold monitor on
error_rateat the hour grain, a forecast-band or period-change monitor onpv, and a collection-gap monitor for minutes with no data. Use the dry run first; it does not send anything and tells you how many times the last 30 days would have fired, which helps you pick a threshold that does not wake anyone for nothing. - Compare before and after. The errors and performance pages have no previous-period comparison, so set one date range, note the numbers, then set the range after the release and compare. Use Open session on an error group to see what the visitor did.
TapCub watches and notifies. It does not approve, block or roll back a release, and a monitor never acts on its alert for you. Those decisions stay in your release process. If a monitor fires, the person on call uses the rules you wrote in the "before" section.
Sources
- UK Ministry of Justice, Avoid code freezes, technical guidance. Checked 2026-11-17.
- Wikipedia, Freeze (software engineering). Checked 2026-11-17.
- Google, Alerting on SLOs, The Site Reliability Workbook (short alert windows and precision). Checked 2026-11-17.
- web.dev, Largest Contentful Paint (LCP) (2.5 s and 4.0 s thresholds at the 75th percentile). Checked 2026-11-17.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Technical Engineer
Focuses on tracking code, SDKs, and API integrations, with practical guidance on implementation checks and troubleshooting.