Fix it now
The time service has found a difference between the local clock and its source that it will not simply apply, so it has stopped trusting itself and stopped serving time to anything else. Kerberos is the next thing to break, because tickets are rejected once client and server disagree by more than the policy tolerance.
w32tm /query /status
w32tm /query /source
w32tm /stripchart /computer:pdc.contoso.com /dataonly /samples:5
w32tm /resync /rediscover
- Read the offset from the stripchart output rather than comparing clocks by eye. It prints the difference against the machine you name, in seconds.
- If this is a virtual domain controller, disable the hypervisor’s time synchronisation integration for it before changing anything else, or the host will keep pulling the clock back.
- Put the machine back on the domain hierarchy if it should be following it:
w32tm /config /syncfromflags:domhier /update, thennet stop w32timeandnet start w32time. - If the clock is still wrong after a resync, set it roughly right by hand and resync again. The service refuses corrections larger than its configured limit, so a very large offset has to be closed manually first.
Fix the top of the hierarchy before the leaves. A member server that disagrees with a domain controller which is itself wrong will look corrected and then drift straight back.
If the status output shows a recent successful synchronisation and a small offset, you are done. The next section explains why a badly wrong clock never fixes itself.
Why it happens
The Windows Time service does not accept every sample it is given. Samples are checked against the phase correction boundaries set by MaxPosPhaseCorrection and MaxNegPhaseCorrection: if a correction is larger than the limit allows, the sample is ignored and the service logs instead of stepping the clock. That is the mechanism behind a machine that stays wrong no matter how many times you run a resync.
The defaults are worth knowing because they explain what you are seeing. On domain controllers both values default to 48 hours. On domain members the default is 0xFFFFFFFF, which means always make the correction, and on stand-alone clients and servers it is 54,000 seconds, fifteen hours. So the same offset can be corrected silently on one machine and refused on another, depending on what it is.
Kerberos is separate from all of this, and it is the reason the event matters. The Kerberos policy setting Maximum tolerance for computer clock synchronization allows a five-minute difference by default between the KDC and the target, and what matters is the relative difference rather than whether either clock is accurate. Once you cross it, authentication fails with a skew error, and on domain controllers replication fails with it too.
A hypervisor and the domain hierarchy are both setting the clock
You have this one if The machine is virtual, the time changes without explanation, and the source in w32tm /query /source is a host integration provider rather than a domain controller.
- Disable the time synchronisation integration service for that guest on the host.
- Restart the guest and confirm the source is a domain controller or a configured peer.
- On a virtual domain controller this is Microsoft’s documented guidance, because a jump in system time can leave lingering objects in caches and stop replication.
The offset is larger than the service will correct
You have this one if Resync appears to succeed and the clock does not move, or the status shows the machine is not synchronised despite a reachable source.
- Read MaxPosPhaseCorrection and MaxNegPhaseCorrection under
HKLM\SYSTEM\CurrentControlSet\Services\W32Time\Config. - Set the clock roughly right by hand so the remaining offset is inside the limit, then run
w32tm /resync. - Decide deliberately what those limits should be afterwards. Microsoft recommends a value other than the maximum, so a machine that comes back with a wild clock cannot drag others with it.
The machine is following the wrong source
You have this one if w32tm /query /source names something unexpected – a manual peer, a stale server, or a local clock – on a machine that should be following the domain.
- Put it back on the hierarchy:
w32tm /config /syncfromflags:domhier /update, then restart the service. - Check nothing sets a manual peer through Group Policy, which will simply put it back at the next refresh.
- Confirm with
w32tm /query /configurationthat Type reads NT5DS on a domain member.
The source itself is wrong
You have this one if Several machines report large offsets at once, and they all follow the same domain controller.
- Check the whole set with
w32tm /monitor, which reports the offsets of the domain controllers in the domain you run it in. - Work up the hierarchy to the forest root PDC emulator and fix its configuration first.
- Then resync downwards. Correcting members against a wrong source achieves nothing.
Run w32tm /monitor in each domain if you have more than one, because it only checks domain controllers in the domain of the computer you run it from.
Full reference
What a wrong clock breaks, and in what order
| Symptom | Underlying value |
|---|---|
| Kerberos authentication fails with a skew error | 0x25, KRB_AP_ERR_SKEW – the clock skew is too great |
| Windows reports a time difference with the domain controller | 0xC0000133 – the time at the primary domain controller differs from this machine by too large an amount |
| Replicate now in the sites console fails | 1398, ERROR_TIME_SKEW, 0x576 – there is a time and/or date difference between the client and the server |
| A ticket is refused as not yet valid | KRB_AP_ERR_TKT_NYV in a network trace: the ticket’s validity starts later than the target’s clock |
LSASrv Event ID 40960 is where 0xC0000133 usually surfaces. On a domain controller whose replication is failing, look for 40960 entries citing the source domain controller with that extended error – it is the clearest single confirmation that time, not permissions, is the problem.
Measuring rather than guessing
| Command | What it tells you |
|---|---|
w32tm /query /status |
Source, stratum, last successful sync time, and the current offset |
w32tm /query /source |
Just the source, which is the fastest way to spot a hypervisor provider |
w32tm /stripchart /computer:<target> /dataonly /samples:5 |
The offset against a named machine, in seconds, sampled repeatedly |
w32tm /monitor |
Offsets across the domain controllers in this domain |
dcdiag /test:CheckSecurityError |
Reports excessive time skew among domain controllers as part of a wider security check |
The phase correction limits, in full
| Machine type | Default MaxPos / MaxNegPhaseCorrection |
|---|---|
| Domain controller | 172,800 seconds – 48 hours |
| Domain member | 0xFFFFFFFF – always make the correction |
| Stand-alone client or server | 54,000 seconds – 15 hours |
Microsoft’s recommendation is to set these to something other than the maximum, precisely so that a machine which acquires a wild clock – restored from a snapshot, or briefly pointed at a bad source – cannot pass it on. If you impose limits on the domain controllers, a single domain controller with a large jump is quarantined by its neighbours rather than followed.
Doing it in the right order
- Fix the forest root PDC emulator first, so the hierarchy has a correct root.
- Fix the other domain controllers, checking with
w32tm /monitorin each domain. - Disable hypervisor time synchronisation on any virtual domain controller before you correct it.
- Then correct members, which will mostly correct themselves once their source is right.
- Re-check after a full day. Time problems that come back are configuration problems; time problems that stay fixed were events.
Kerberos tolerance is a policy, not a constant
The five-minute figure comes from the Kerberos policy setting Maximum tolerance for computer clock synchronization in the default domain policy. It is adjustable, and it is not what the time service uses for its own decisions. Two consequences follow: raising it to paper over a drifting estate weakens the replay protection it exists to provide, and lowering it makes authentication fail sooner than any time service event will warn you about.
Every code this article covers
| Code | What it points at | Source |
|---|---|---|
Event ID 50 |
Written by the time service when it finds a difference it will not accept and stops synchronising. No Microsoft page publishing this ID’s text was found, so confirm the state with w32tm /query /status rather than reading the number alone |
not published by the vendor |
0x25 |
KRB_AP_ERR_SKEW: the clock skew is too great, measured against the Kerberos policy setting Maximum tolerance for computer clock synchronization | Microsoft Learn |
0xC0000133 |
The time at the primary domain controller differs from the time at this machine by too large an amount; commonly seen as the extended error inside LSASrv Event ID 40960 | Microsoft Learn |
1398 |
ERROR_TIME_SKEW, 0x576: there is a time and/or date difference between the client and the server | Microsoft Learn |
Confirm the fix worked
w32tm /query /statusshows a recent successful synchronisation and a small offset.w32tm /stripchartagainst the PDC emulator shows offsets in fractions of a second.w32tm /monitorshows the domain controllers agreeing with each other.dcdiag /test:CheckSecurityErrorno longer reports excessive skew.- No new LSASrv 40960 events citing 0xC0000133, and replication runs clean.
Questions people ask about this
Why does the clock not correct itself when a source is available?
Because the correction is larger than MaxPosPhaseCorrection or MaxNegPhaseCorrection allows, so the sample is ignored and an event is logged instead. Set the clock roughly right by hand, then resync.
Is five minutes a hard limit?
It is the default of the Kerberos policy setting Maximum tolerance for computer clock synchronization, and it is adjustable. What matters to Kerberos is the relative difference between the KDC and the target, not whether either clock is accurate.
Should I disable hypervisor time synchronisation on every virtual machine?
On virtual domain controllers, Microsoft’s guidance is to disable it so the guest follows the domain hierarchy and is protected from a jump when restored from a saved state. For ordinary member virtual machines it matters less, but one source is still better than two.
Which machine do I fix first?
The top of the hierarchy. Correcting a member against a domain controller that is itself wrong just moves the error around.
Does any of this need a purchase?
No. Every step here is configuration or a registry value.
