Time zones, clocks and traffic peaks you can see coming
Updated Nov 10, 2026 · 4 min read
Some traffic spikes are a surprise. Others are on the calendar, down to the second. New Year’s Eve is the clearest case. Everyone opens the same app at midnight, but midnight doesn’t arrive everywhere at once. It arrives one time zone at a time, so a global app sees a string of sharp spikes through the day rather than one giant one. Planning for that takes some care with time zones, with clocks, and with when you add capacity.
Midnight moves west
The larger a zone’s offset from UTC, the earlier its midnight. The first places to reach the new year sit in the Pacific, and the last are in the Americas and the Pacific islands east of the date line. Some zones are 30 or 45 minutes off the hour, so not every spike lands on the hour.
| City | Offset in December | Midnight in UTC |
|---|---|---|
| Auckland | UTC+13 | 11:00 on Dec 31 |
| Sydney | UTC+11 | 13:00 on Dec 31 |
| Tokyo | UTC+9 | 15:00 on Dec 31 |
| New Delhi | UTC+5:30 | 18:30 on Dec 31 |
| London | UTC+0 | 00:00 on Jan 1 |
| New York | UTC−5 | 05:00 on Jan 1 |
Size for the biggest zone
Because each zone peaks on its own, the number to plan for is the biggest zone’s spike, not all your users at once. If 30% of 20 million users live in one zone and a third of them open the app in the first minute after midnight, that’s 2 million opens in 60 seconds, or about 33,000 a second. It fades within the hour, then the next zone arrives.
Store zone names, not offsets
An offset like +10:00 is only true for part of the year. Sydney is UTC+10 in July and UTC+11 in December, because of daylight saving. Save +10:00 in July, schedule a midnight push with it in December, and it lands at 1 AM. Store the zone’s name from the IANA time zone database, like Australia/Sydney, and work out the offset at the moment you need it. Keep that database up to date, since countries change their rules with little notice.
Whose clock is right
Servers keep time with NTP, usually to within a few milliseconds. Phones and laptops are often close, but some are minutes off. A countdown that trusts the device clock hits zero at the wrong moment. The fix is to ask the server for its time once, measure how long the request took, and assume the answer is half a round trip old.
sent = Date.now()
serverTime = await fetch('/time')
received = Date.now()
offset = serverTime + (received - sent) / 2 - received
// From now on, the true time is Date.now() + offsetOne request gives an offset good to a fraction of the round trip, and the page then counts down on its own clock plus that offset. Asking every second adds load and doesn’t make it more accurate.
Serve what you can statically
A countdown page is the same for everyone, and the browser knows its own time zone, so the browser can do the counting. Serve the page from a CDN and your servers barely notice that millions have it open. Save the servers for what truly changes at midnight, the greetings people send.
Pre-scale for spikes you know about
Autoscaling reacts to load that is already here, and a new server takes minutes to be ready. A spike that arrives in seconds is over the edge before the first new server boots. When you know the moment, scale on a schedule instead, an hour before each big zone’s midnight, and back down after.
- Weeks before, load test at more than the biggest expected spike.
- Days before, freeze deploys so nothing new ships into the busiest night.
- An hour before each big zone, scale up and warm caches, since a fresh server with an empty cache is slow.
- During the spike, switch off work that can wait, like recommendations or analytics.
- After, scale back down so you only pay for the hour you needed.
Scheduling a message per zone
Sending “Happy New Year” to every user at their own midnight is a scheduling job, not a loop at midnight UTC. Group users by zone name, and a few minutes before each zone’s midnight, enqueue that zone’s messages with a time to deliver. Workers drain the queue at a rate your push provider accepts. A zone with tens of millions of users still takes a few minutes to send, so start early enough that the first arrive on time.