Nobody wants to spend Thanksgiving dinner on an incident bridge. Nobody wants their customers there either. That is a reasonable argument for slowing production changes during the holidays. Fewer people are available, the engineers who understand a particular failure may be away, and some customers are entering their busiest weeks of the year.
But “we stopped deploying” is not the same as “we made production safer.” A freeze can protect a critical period. It can also leave known problems running, encourage a last-minute release rush, and build a queue that everyone wants to empty in January.
The question is not whether freezes are good or bad. It is whether the risk of making a change is greater than the risk of waiting.
Why a pause can make sense
The same failure can have very different consequences depending on when it happens. A checkout problem during an ordinary Tuesday is bad. During Black Friday, it can cost a merchant a meaningful part of their annual business. A release that needs a specialist to recover is more dangerous when that specialist is on leave. Automation helps, but it does not remove that human dependency.
AWS explains this trade-off in its deployment practices. Its deployment windows exclude nights, office holidays, and weekends because some failures still need manual intervention. It also warns that overly narrow windows allow changes to accumulate, making the next deployment larger and potentially riskier. Both sides matter. AWS Builders’ Library
There is another reason to pause that engineers sometimes overlook: customers need predictability. In its 2018 Black Friday preparation write-up, Shopify described freezing features several weeks before the event so merchants would not have to learn new workflows during their busiest period. A shorter code freeze followed, allowing critical fixes. That was not simply fear of deployment. It was consideration for the people using the product. Shopify Engineering
Protecting a business-critical week is a valid engineering decision. Treating every change as equally dangerous is not.
What the hyperscalers actually tell us
There is no single publicly documented holiday calendar shared by AWS, Google Cloud, Azure, and OCI. Their public material describes a more useful principle: control when changes happen, how far they spread, and what evidence allows them to continue.
Google Cloud describes progressive deployment waves. A wave that extends into a night or weekend may finish, but the next wave waits for the managing team’s business hours. Health checks and observation periods determine whether rollout continues. That is a deliberate pause, not an assumption that automation makes staffing irrelevant. Google Cloud’s approach to change
Azure describes staged deployments, health signals, and observation periods, with exceptional acceleration for urgent situations such as security fixes. Its published guidance does not establish a company-wide holiday freeze calendar. Azure safe deployment practices
Oracle’s Alloy documentation explicitly supports operator-requested freezes, while recommending they remain short. The stated reasons are practical: freezes delay security fixes, improvements, and backlog clearance. This is an Alloy policy, not proof of a universal OCI holiday schedule. Oracle Alloy change management
The lesson is not “copy what a hyperscaler does.” It is to understand what the restriction is protecting.
The trouble can start before the freeze
Announce a deadline and people try to beat it. A feature that could have waited suddenly needs to ship Thursday because the freeze starts Friday. Testing gets compressed. Observation time disappears. Several teams make their final changes together. The calendar has unintentionally created the behaviour the freeze was supposed to prevent.
OneSignal documented a concrete example from its 2020 election-related freeze. A billing feature deployed shortly before the freeze caused customer problems during the protected period. The team initially held the fix, then made an exception. Its retrospective proposed a buffer between the feature cutoff and the hard freeze. OneSignal’s code-freeze experience
That is a useful distinction: the last permitted deployment should not happen at the last moment anyone is available to watch it.
If a change needs two days of observation, the deadline must leave those two days intact. The freeze date is not permission to shorten qualification.
Production does not stop changing because we stop deploying
Traffic keeps moving. Data grows. Certificates approach expiry. Hardware fails. Dependencies change. Security vulnerabilities are discovered. A freeze preserves the deployed state. It does not preserve the conditions around it.
Imagine a storage service with a known capacity problem. Blocking an expansion may avoid the immediate risk of changing the fleet, but it leaves the service closer to exhaustion. Neither choice is risk-free. The same applies to a patch for an actively exploited vulnerability or a repair that restores redundancy.
This is why “emergency changes only” needs a definition before the holiday begins. Who can approve one? What evidence is required? Who will validate it? Who can recover the service if it goes wrong?
An exception should not mean abandoning engineering discipline. It should mean that the risk of waiting has become greater than the risk of acting.
January is part of the decision
A quiet December does not tell the whole story. What was postponed? Are the changes still independently releasable? Have their assumptions changed? Will multiple teams restart against shared infrastructure at the same time?
DORA’s guidance favours small batches because they shorten feedback and make problems easier to identify. Accumulating work and releasing it together gives up those advantages. That does not prove holiday freezes cause January outages. It does explain why an uncontrolled restart deserves scrutiny. DORA: working in small batches
The end of a freeze should reopen the normal release process, not create a backlog-clearing event. Recheck dependencies. Revalidate older artifacts where needed. Release small batches. Keep the usual observation periods. Being delayed does not make a change more urgent or better tested.
What I would expect from a useful freeze
A holiday freeze should answer five questions clearly:
- What are we protecting? A customer peak, reduced staffing, a financial close, or a specific critical event. Name the reason.
- What is actually restricted? Define the services and change types. A broad network migration and an isolated documentation update do not carry the same risk.
- What must be ready beforehand? Capacity, certificate validity, recovery procedures, on-call coverage, and enough observation time after the final releases.
- What can still proceed? Define security, reliability, and incident-recovery exceptions, with named decision-makers and available responders.
- How do we restart? Agree on sequencing and validation before the queue starts growing.
Then review the whole period: the rush before, the exceptions during, and the releases afterward. Count delayed fixes and customer impact, not just incidents inside the freeze window.
Are we postponing the inevitable?
Not necessarily. Failure is not inevitable, and timing matters. Moving an optional change out of a high-stakes week can genuinely reduce harm.
But waiting does not repair a weak release process. Nor does a quiet dashboard prove that every blocked change was dangerous.
My position is simple: use short, clearly scoped freezes when the business or staffing situation justifies them. Keep a disciplined path for changes that make production safer. Plan the restart as carefully as the pause.
Let people take their holidays. Protect customers when it matters most. Just do not confuse a closed deployment window with a healthy service.