On October 15, 2026, delegates to the General Conference on Weights and Measures voted in Versailles to stop adding leap seconds to Coordinated Universal Time. From May 20, 2027, UTC will simply keep counting, and the gap between atomic time and the Earth's rotation will be allowed to grow to a full hour before anyone has to touch it again. That sounds like a footnote for astronomers until you trace where your servers actually get their time: as a recent analysis of how one clock in the sky keeps the modern economy running lays out, exchanges, carriers and grid operators all lean on the atomic clocks aboard GPS satellites, and the GPS time scale has never contained a single leap second. It has run without interruption since January 6, 1980, which leaves it 18 seconds ahead of UTC, so any receiver that outputs UTC subtracts a broadcast offset before the result flows down to NTP, to your kernel and eventually to time.Now(). In six months that offset freezes for good, and a whole family of bugs that lived in the gap between those two clocks loses its habitat.
The Second That Kept Breaking Production
Leap seconds exist because the planet is a mediocre clock. Since 1972, whenever the Earth's rotation drifted toward 0.9 seconds away from atomic-based UTC, an extra second was scheduled, always at the end of June or December. Twenty-seven were inserted, every one of them positive, on top of a 10-second offset baked in at the start, which is why UTC now trails International Atomic Time by 37 seconds.
Software never liked them. POSIX time insists that every day contains exactly 86,400 seconds, so a Unix timestamp has no way to say 23:59:60. Many systems coped by stepping the clock back, which means 23:59:59 happens twice and wall-clock time briefly runs backwards.
On June 30, 2012, that backwards step met a Linux kernel bug. The high-resolution timer subsystem was never told the clock had been set, futex waits began timing out the instant they started, and Java and MySQL workloads spun their CPUs at full load. Reddit, LinkedIn, Mozilla and Yelp all had a rough evening, and the Amadeus Altéa reservation platform went offline for about an hour, sending Qantas and Virgin Australia staff back to checking passengers in by hand.
The leap second at the end of 2016 found a different layer of the stack. Cloudflare's postmortem of its New Year's Day DNS failure traces the outage to one assumption inside its Go-based RRDNS service: that time never goes backwards. Upstream round-trip times were measured with time.Now(), which in those days read only the wall clock, so when the clock was reset some measurements came out negative. The smoothed average followed them below zero and eventually landed in rand.Int63n(), which panics on a negative argument. At peak about 0.2 percent of all DNS queries were affected, every one of them for customers using CNAME records, and a fix was live worldwide by 06:45 UTC.
The more lasting result was a change to the language itself. Russ Cox's proposal for monotonic elapsed-time measurement cited the Cloudflare incident, and since Go 1.9 shipped in August 2017, every time.Time returned by time.Now() carries a hidden monotonic reading that comparisons and subtractions use automatically. That fix has outlived its trigger in a fitting way, because the 2016 leap second is now all but certain to be the last one ever applied.
Why the Deadline Jumped From 2035 to 2027
The same conference had already agreed in 2022 to retire the leap second "in, or before, 2035." Then the planet forced the schedule. The Earth has been spinning slightly faster for several years, and a 2024 study in Nature by geophysicist Duncan Agnew projected the first-ever negative leap second, a minute containing only 59 seconds, around 2029; without melting polar ice slowing the rotation, it would have been due in 2026. A March 2025 workshop of timing and Earth-rotation experts then put the probability of a negative leap second at 30 percent by 2035.
That number is the real story. Plenty of software has at least been exercised against an inserted second, but almost nothing has been tested against a deleted one, because it has never happened. When the BIPM put that figure to users and industry stakeholders, the answer, according to Scientific American's report on the leap-hour proposal, was that even a 10 percent chance was too much.
Resolution 5 of the 28th CGPM therefore fixes a date and a limit: continuous UTC from May 20, 2027, which is World Metrology Day, and a maximum gap of 3,600 seconds between UTC and UT1, enough to last several centuries. The text is unusually blunt about the motivation. It compares the investment a negative leap second would demand to the preparations for the millennium bug, notes that the inconsistent ways global service providers absorb leap seconds produce time that does not conform to the SI, and warns that further delay could push users to drop UTC in favor of GNSS time. Read between the lines and the official time scale changed partly to avoid being replaced by the clock in the sky.
Smearing, the Workaround You Can Finally Retire
The largest platforms stopped trusting the leap second long ago. Google has smeared leap seconds since 2008, and its current standard stretches the extra second linearly across 24 hours from noon to noon UTC, so each second runs about 11.6 parts per million long. AWS adopted the same smear for its NTP endpoints. Meta went its own way with a 17-hour quadratic smear starting at midnight UTC. During the window, a clock following Google's recipe sits up to half a second away from true UTC, and smears built on different recipes disagree with each other, which is why AWS documentation warns against mixing smeared and unsmeared sources in one client configuration.
There is a quieter trap even inside a single cloud. On EC2, the NTP service smears, but the PTP hardware clock available on supported instances does not, and it would apply a leap second the standard way. Same provider, same atomic reference, two different answers during a leap second.
After May 2027 all of that goes inert. There will be no leap warnings in NTP packets, no smear windows and no second number 60. Nothing needs to be ripped out on day one, since the code will simply never fire, but it deserves a comment marking it as legacy, so nobody burns an hour of a future incident wondering what leapsecmode was protecting.
What Continuous UTC Does Not Fix
Ending the leap second deletes a discontinuity. It does not delete the reasons clocks disagree, and in a few places it moves complexity rather than removing it.
The Earth offset becomes your problem instead of UTC's. UT1 minus UTC used to be guaranteed to stay under 0.9 seconds, so plenty of code treated it as noise. From 2027 it is allowed to drift past a second and keep going. If you compute sidereal time, satellite passes, antenna or telescope pointing, or anything else tied to the planet's real orientation, you can no longer treat UTC as a stand-in for UT1; you need actual Earth-orientation data from the IERS, and every assertion that the correction stays below one second is a slow fuse. Even radio time signals were designed around the old limit, which is why the ITU gave radio time services a transition period that runs to 2035, or to 2040 where equipment cannot be replaced sooner.
History still contains 61-second minutes. Logs, market data and regulatory archives from 2016 and earlier can contain 23:59:60, and Unix timestamps silently skip the 27 real seconds inserted between 1972 and 2017. Parsers must keep accepting what your systems will never produce again.
The translation layer can still break. In January 2016, right after operators retired SVN 23, the oldest satellite in the constellation, a faulty ground-software upload left 15 GPS satellites broadcasting a UTC correction that was off by about 13 microseconds. The atomic clocks were fine and navigation worked normally, yet telecom and digital-radio timing gear raised alarms for the better part of two days. Freezing the leap-second count does nothing to the rest of that correction message, so continuous UTC can still arrive with a wrong number attached.
2038 remains a double feature. On January 19, signed 32-bit time_t overflows. On November 20, receivers that decode only the legacy GPS navigation message hit the third 10-bit week-number rollover, the same failure mode that took down New York City's municipal wireless network in April 2019.
The Monotonic Rule Survives Everything
If one habit should outlive the leap second, it is the one Cloudflare learned the hard way: wall clocks label events, monotonic clocks measure them. NTP still steps clocks after long outages, virtual machines still wake from suspension with stale time, and somebody will still run date -s on a production box at 3 a.m. Continuous UTC changes none of that.
Go makes the boundary unusually visible, because the monotonic reading lives inside time.Time and quietly vanishes the moment the value leaves the process:
start := time.Now() // wall clock + monotonic reading
doWork()
elapsed := time.Since(start) // monotonic: immune to clock steps
// The monotonic reading means nothing outside this process,
// so every serializer drops it.
b, _ := start.MarshalJSON()
var restored time.Time
_ = restored.UnmarshalJSON(b)
lag := time.Since(restored) // plain wall-clock math again
The same split exists everywhere: time.monotonic() in Python, performance.now() in browsers and Node, System.nanoTime() in Java, CLOCK_MONOTONIC in C. In Go the fallback to wall-clock math happens at the process boundary; in every language it happens at the machine boundary, where subtracting two timestamps is only as good as the agreement between two clocks. That is where you need timestamps that carry an explicit error bound (AWS's open-source ClockBound daemon exists for exactly this) rather than arithmetic and hope.
A Cleanup List for the Next Six Months
Six months is plenty of time to make May 20 a non-event. A practical pass looks like this:
-
Inventory your leap-second handling. Search configs and code for
leapsecmode,leapfile,leap-seconds.list, smear settings and any special case for second 60, then label it legacy rather than deleting it. -
Keep
23:59:60parseable, stop generating it. Historical records need the parser; no future record will ever need the generator. - Audit durations, timeouts and rate limiters for wall-clock subtraction, especially anywhere timestamps are persisted or sent over the network.
- Treat UT1 as data, not a constant. Pull Earth-orientation values from IERS products through your astronomy or orbital library and remove hard-coded sub-second bounds.
-
Keep tzdata fresh in your images. The
leap-seconds.listshipping today expires on June 28, 2027, and ntpd logs a warning once a leap file expires, so an image built today will start complaining about five weeks after continuous UTC begins unless it picks up a refreshed copy. -
Put 2038 on the risk register for embedded devices, older GPS receivers and anything still compiled with a 32-bit
time_t. - Keep alerting on clock offset. The scheduled discontinuity is going away; drift, jamming and bad upstream data are not.
The Rare Bug Fixed by Deleting the Feature
Engineers seldom get to watch a standards body close an entire class of bugs by removing the feature that caused it, and that is what happened in Versailles. After May 20, 2027, GPS time and UTC will sit 18 whole seconds apart for the foreseeable future, smears will never run again, and the 59-second minute nobody ever tested will stay hypothetical. The same meeting also adopted resolutions on the future definition of the second and on a reference time scale for the Moon, so timekeeping is far from done changing. For most codebases, though, the remaining work is small and finite: clean up what the leap second left behind, keep measuring durations with the right clock, and remember that the "UTC" arriving from orbit is a number somebody computed, not a law of physics.
Top comments (0)