TL;DR
- ▸Allan deviation measures short-term frequency stability across a time window. Rising Allan deviation at τ=100s is the classic signature of an OCXO nearing end-of-life — and it is invisible to threshold-based offset alerting.
- ▸Sync Insight computes and streams Allan deviation continuously per host. A Grafana panel overlaying the current week against the same week a year ago shows aging trends over a multi-month horizon.
- ▸The operational value is that oscillator replacement becomes planned maintenance rather than incident response. A year-long aging trend gives 6+ weeks of runway to schedule a refresh before the oscillator becomes a compliance or operational risk.
What Allan deviation actually measures
Allan deviation, written σ_y(τ), is a measure of frequency stability at a given averaging time τ. Low Allan deviation at a given τ means the oscillator's frequency is stable when averaged over that window; high Allan deviation means it is noisy. Plotting σ_y(τ) across a range of τ values — typically from sub-second to hundreds of seconds — produces the Allan deviation curve, which is the characteristic fingerprint of the oscillator's behaviour.
Different oscillator types produce characteristically different curves. A good-quality OCXO at τ=1s might be around 3×10⁻¹¹; at τ=100s it drops into the 10⁻¹² region as averaging smooths short-term noise; at τ=10,000s it starts rising again as long-term drift (aging, temperature) dominates. A Rubidium Black+ sits approximately an order of magnitude below at the same τ values — 1.5×10⁻¹¹ at τ=1s, and sub-10⁻¹² at τ=100s. Caesium is lower still. White Rabbit over fibre is a different kind of curve entirely — sub-nanosecond continuous, essentially flat across τ.
The metric matters operationally because it captures information that raw offset measurements do not. A host's offset to its upstream reference can be stable and small (say, consistently within 50 ns) while the Allan deviation at τ=100s is slowly climbing month-on-month. That means the oscillator is still producing a usable clock, but its short-term stability is degrading — and six weeks or six months later, the same curve will cross the threshold at which the offset will start to matter. Allan deviation is the leading indicator; offset is the lagging indicator.
Why τ=100s is the specific number
For most operational oscillators, Allan deviation at τ=100s captures the transition between short-term white frequency noise (dominated by quartz resonator quality) and long-term flicker / random-walk noise (dominated by aging and environmental sensitivity). Aging effects start to show up first in the τ=100s to τ=1,000s range, weeks to months before they become visible in raw offset measurements.
Why threshold-based alerting misses aging
Most timing monitoring platforms alert when the offset exceeds a threshold. That is the right thing to alert on for acute events — a GNSS failure, a PTP path disruption, a downstream client rejecting the offered time — because those events produce a fast, clear excursion in the offset measurement. A one-second offset spike above 100 µs rings a bell, and the on-call engineer responds.
Aging does not look like that. An OCXO slowly drifting out of spec over 18 months produces no threshold violations along the way; the offset stays within tolerance until the day it does not, at which point the alerting fires when the oscillator has already aged past its useful operating range. The on-call engineer is responding to a preventable event that should have been scheduled maintenance three months ago.
Trend-based alerting is the antidote. Instead of 'alert when Allan deviation at τ=100s exceeds X', the useful alert is 'alert when Allan deviation at τ=100s has increased by Y% over a rolling 90-day window'. That catches the gradient rather than the absolute value, and it fires when the aging trend is still in its early phase — when replacement can be scheduled during a maintenance window rather than triggered by a P1 incident.
How Sync Insight surfaces the curve
The TimeBeat Agent computes Allan deviation continuously for the host clock at several standard τ values — typically τ=1s, τ=10s, τ=100s and τ=1,000s. The computation is cheap (the agent already has the sample data from its own steering cycle) and the output is streamed as part of the standard 167-field telemetry output.
Sync Insight's Elasticsearch index stores the Allan deviation values alongside every other clock-state measurement at 1-second resolution. For deployments on the 90-day retention tier, this means three months of high-resolution Allan deviation history is available for querying. For deployments on 12-month retention (Enterprise's optional add-on), the history supports year-on-year trending — the most useful form of aging analysis because year-on-year eliminates seasonal environmental factors that otherwise confound the trend.
The Grafana panel configuration that matters most is the 'Allan deviation at τ=100s, 90-day trend, with previous-year overlay'. Two traces: the current 90-day curve, and the 90-day curve from exactly one year ago. A healthy oscillator has the two curves overlapping; an aging oscillator has the current trace consistently above the previous-year trace, with the gap widening. The panel tells you, at a glance, whether the oscillator is aging faster than expected.
What an aging OCXO actually looks like
A typical OCXO aging trace unfolds over 2–5 years. In the first year after commissioning, Allan deviation at τ=100s is stable around the datasheet specification — say, 5×10⁻¹² for a good-quality OCXO. In year two, the trace starts to climb slightly, maybe 10–15% above the year-one baseline; this is the first warning sign, invisible to threshold-based alerting but clear in the year-on-year overlay.
By year three or four, the trace has doubled. Raw offset is still within tolerance because the OCXO's aging compensation is still pulling it back towards the reference — but the servo is working harder, the short-term stability is noticeably worse, and the unit's behaviour under GNSS-denial scenarios is degrading. A holdover test at this point would show drift that is 2–3× the datasheet specification, which is the empirical reality of most OCXO deployments that have not been maintained.
By year five or six, the OCXO starts producing observable offset excursions under normal operation. By the time the threshold-based monitoring catches this, the unit has already been in a degraded state for 18–24 months. The aging trace had been visible in Sync Insight the entire time, but no one had been looking at it because the offset was still within tolerance. This is the scenario that Allan deviation observability is designed to prevent.
What you do with the information
The practical output of Allan deviation observability is maintenance planning. A year-on-year trend showing persistent 20–30% degradation at τ=100s is the signal to schedule oscillator replacement during the next maintenance window. The hardware is still fit for purpose now; it will not be fit for purpose in six months. Acting on the leading indicator converts oscillator replacement from unplanned incident response into planned preventive maintenance.
Rubidium aging — a different signature
Rubidium oscillators age differently from OCXOs. Rubidium's long-term drift is typically much lower — the published long-term drift for the Open Time Appliance's Rubidium Black+ is <60 µs per year — and the Allan deviation at τ=100s stays near the datasheet specification across the oscillator's useful life. What degrades on a Rubidium oscillator is typically the lamp — the physical component that generates the Rubidium resonance signal — and the degradation is usually abrupt rather than gradual.
For Rubidium deployments, the operationally useful observability is less about Allan deviation trending and more about monitoring the internal health telemetry. Sync Insight exposes the Timecard power fields — bus voltage, shunt voltage, current draw, power consumption — and a unit whose current draw is slowly climbing over months is typically in the early stages of lamp degradation. The refresh cycle for a Rubidium lamp is 10–15 years under normal conditions; the Sync Insight telemetry will surface the need for a refresh 3–6 months before the unit becomes operationally unreliable.
Caesium oscillators have their own characteristic aging signature — Caesium beam depletion over 15–25 years — that is captured through different telemetry fields. Sync Insight's approach is consistent across oscillator types: surface the leading indicator for each oscillator class, alert on the trend rather than the threshold, convert replacement from incident to maintenance.
Wiring up the Grafana panel
The practical configuration for most deployments is a single Grafana row dedicated to Allan deviation trending. Four panels in that row capture the full picture: Allan deviation at τ=1s (short-term noise floor), τ=10s (servo operating range), τ=100s (aging indicator), and τ=1,000s (long-term drift). Each panel shows the current 90-day trace. For deployments with a year or more of retention, overlaying the year-ago trace on each panel is the single most valuable addition.
Alerting configuration is simpler. One alert per host: 'Allan deviation at τ=100s has increased by >25% over the rolling 90-day average compared to the 90-day average 9–12 months ago'. This fires when the aging trend is established and still pre-critical. For Rubidium deployments, add an alert on current-draw trend (>15% increase over rolling 90 days vs year-ago baseline) as the equivalent leading indicator for lamp degradation.
The alerts should fire to a maintenance-planning channel, not a P1 on-call rotation. The whole point of trend-based alerting on aging is that there is no urgency — weeks of runway to plan the refresh. Routing these alerts into the Slack channel where maintenance windows are planned, or into the ticketing system where the infrastructure team tracks scheduled work, is the right operational treatment.
The procurement case
Allan deviation observability is part of every Sync Insight tier — Professional, Enterprise and PAYG all produce the same telemetry. A team on PAYG running a 10-device trial can configure the same Grafana panel as an Enterprise-tier deployment. The difference between tiers is retention: longer retention enables longer trend windows, which makes year-on-year overlays possible.
For deployments focused on oscillator aging observability, the retention decision is the relevant one. 30 days on Professional is enough to see an active aging event in progress but not enough to catch it early. 90 days on Enterprise (or on Professional with the 90-day add-on) catches most aging trends in time for planned maintenance. 12 months on Enterprise+ (via the +£500/month add-on) enables year-on-year overlays, which for long-running oscillator populations is the gold-standard visibility.
The economic case is the cost of planned vs unplanned replacement. A planned oscillator refresh during a maintenance window is typically 2–4 hours of engineering time plus the hardware cost. An unplanned replacement triggered by a P1 incident is typically the same hardware cost plus incident-response overhead, plus whatever operational impact the degraded clock caused before the replacement. Across a multi-year deployment with a population of oscillators approaching end-of-life concurrently, the difference is material.
Start with a trial
A 30-day Sync Insight trial on PAYG at £1.12/device/day surfaces the aging telemetry for your existing oscillators within hours. The trial gives you one month of trend data — usually enough to see which of your units is actively aging. The year-on-year overlay comes later, but the first-month trial tells you what to look at.
Frequently asked questions
Does the agent need special configuration to compute Allan deviation?+
What if we only have a short deployment history?+
Is Allan deviation a good alert trigger or only a dashboard panel?+
Does Allan deviation work for NTP-only deployments without hardware timestamping?+
Related reading
Blog · Pricing
The New Sync Insight Pricing Lanes — What Auto-Cap Means For Your Deployment
Four pricing lanes from £1.12 PAYG trials through to bespoke Enterprise+ contracts. The overage maths, why the 167-device auto-cap eliminates the usual procurement step-change, and which lane actually fits your deployment right now.
Blog · Compliance
Monitoring A Clock vs Proving A Clock: MiFID II, DORA, CAT
The difference between observability and attestation — why the direction of travel in ESMA, FCA and SEC guidance is making log-file compliance structurally insufficient, and what independent verifiability means in practice for a regulator's audit team.