How the adaptive threshold is calculated
The automatic mode of response-time monitoring in the manual works with two quantities that come about at different times. The first is the baseline. Once a night, XICTRAQ® calculates the mean μ and the standard deviation σ of the measured response times for every website and each of the 168 hours of a week. It uses the last 14 days, and an hour of week becomes usable from 20 data points (XICTRAQ manual). Only successful checks of the start URL count. Outages, subpages, port checks and manual scans stay out.
The second quantity is the current value. A check runs every 5 minutes (XICTRAQ manual): it collects the response times from a rolling window, sorts them and determines the 95th percentile, P95 for short. You set the window length per website; the default is 15 minutes, and anything from 5 to 180 minutes is possible (XICTRAQ manual). The values come from uptime monitoring, which checks every website from three European locations (XICTRAQ manual). If the window holds fewer than five values, no decision is made.
threshold = max(μ + 3σ ; 1.5 · μ ; 500 ms)
μ mean response time for this hour of week, last 14 days
σ standard deviation of the same values
P95 95th percentile of the values in the rolling window
alert as soon as P95 >= threshold
no usable baseline: fixed threshold, otherwise 5,000 msEach of the three parts has its own job. The distance of three standard deviations is the classic control limit of statistical process control. One and a half times the mean applies when the values of an hour are so close together that three standard deviations amount to less than half the mean. The 500 milliseconds are an absolute floor for fast sites, where a jump from 150 to 300 milliseconds stands out statistically but is rarely a problem for visitors. The 5,000-millisecond fallback mainly affects new websites until their baseline is in place.
| Website situation | Which threshold applies |
|---|---|
| Fixed threshold entered | the fixed threshold, even with automatic mode switched on |
| Automatic mode on, baseline with at least 20 values and measurable spread | max(μ + 3σ; 1.5 · μ; 500 ms) |
| Automatic mode on, baseline missing or too thin | 5,000 ms as a coarse fallback |
| Fewer than five values in the window | no decision, an open alert closes without an all-clear |
Three sigma and why the floors are needed
The limit at three standard deviations comes from quality control. The e-Handbook of Statistical Methods published by the National Institute of Standards and Technology (NIST) describes it as the usual choice: usually the multiple is 3, hence the name 3-sigma limits (NIST/SEMATECH e-Handbook). The idea behind it: without a special cause, a point should only fall above the upper limit with a probability of one in a thousand (NIST/SEMATECH e-Handbook). For normally distributed values, such a limit lies very close to three standard deviations.
Response times, however, are not normally distributed. They have a hard lower bound, because no response arrives in less than zero milliseconds, and a long tail toward the top. The NIST handbook explicitly names measurements with a lower bound as an example of right-skewed data (NIST/SEMATECH e-Handbook). How long that tail can get is described in Google's SRE book: for a service with an average of 100 milliseconds at 1,000 requests per second, 1 percent of requests might easily take 5 seconds (Google SRE Book).
For the threshold this has a consequence that the NIST handbook spells out. If the distribution is skewed to the right, the upper 3-sigma limit falls short of the 0.001 limit (NIST/SEMATECH e-Handbook). The risk of looking for a cause of upward variation that does not exist is then greater than one in a thousand (NIST/SEMATECH e-Handbook). This is exactly where the two fixed floors come in. They keep sites with very steady, fast responses from alerting on every harmless outlier. An example calculation with assumed values shows how the three parts work together:
- Fast, quiet site: μ = 180 ms, σ = 25 ms. μ + 3σ gives 255 ms, one and a half times the mean gives 270 ms. The 500 ms floor applies.
- Steady hour: μ = 420 ms, σ = 60 ms. μ + 3σ gives 600 ms, one and a half times the mean gives 630 ms. 630 ms applies.
- Evening hour with more spread: μ = 420 ms, σ = 80 ms. μ + 3σ gives 660 ms and is above the 630 ms from one and a half times the mean. 660 ms applies.
- New website without history: no usable baseline. The fixed threshold entered applies, otherwise 5,000 ms.
Even ideal data produce false alarms. A control chart with the standard rule, which only signals a point beyond plus or minus three sigma, fires for no reason every 371 points on average (NIST/SEMATECH e-Handbook). Additional pattern rules such as the Western Electric rules make the chart more sensitive to a gradual shift, but raise false alarms to about one in every 91.75 points (NIST/SEMATECH e-Handbook). A check every five minutes means 288 decisions per day. The figures do not carry over one to one to a P95 window, but they still show the order of magnitude.
The amount of data matters too. For setting up a classic control chart, the NIST handbook quotes Shewhart's rule of thumb that at least 25 samples of four values each are needed before a process can be called in control (NIST/SEMATECH e-Handbook). The baseline requires 20 data points per hour of week (XICTRAQ manual). With one check per minute from three locations, two weeks add up to 360 values per hour of week. What matters is something else. All of these values come from exactly two calendar days, because an hour of week occurs only twice in 14 days. Two Friday evenings are the baseline's entire memory of Friday evening.
What the P95 in a short window measures
The P95 is determined by the common rank method: the values are sorted in ascending order, and the P95 is the value at rank 0.95 times the count, rounded up. In small windows this has a consequence worth knowing. With up to 19 values the rank lands on the last place, so the P95 is simply the largest single value in the window. Only from 20 values on does it separate from the maximum.
How many values end up in the window depends on the check interval. The plan sets the shortest interval: Free checks every 5 minutes, Starter and Business every minute, Agency every 30 seconds (XICTRAQ manual). Because three locations check, each interval usually adds three values. The arithmetic gives the following picture:
| Check interval | Window | Values in window | The P95 is |
|---|---|---|---|
| 5 minutes | 5 minutes | 3 | no decision (under 5 values) |
| 5 minutes | 15 minutes | 9 | the largest value |
| 5 minutes | 60 minutes | 36 | the second-largest value |
| 1 minute | 15 minutes | 45 | the third-largest value |
| 30 seconds | 15 minutes | 90 | the fifth-largest value |
For peak week this means: with few values in the window, a single slow request is enough to lift the P95 above the threshold. That makes the alert sensitive and fits a recommendation from the SRE book, according to which the 99th percentile response time over a small window, such as one minute, gives a very early signal of saturation (Google SRE Book). In a week with many short load spikes, the same sensitivity can lead to frequent alerts. Why percentiles say more than the average is explained in the post p95 and p99 instead of averages.
Work out how many values land in your window. Under five, no decision is made; under 20, the slowest single value decides. If you check every five minutes and want a calmer signal, you need a window longer than 30 minutes, which in the selection is the 60-minute option. With one check per minute, the default 15 minutes are usually enough.
Why response time under load often rises in steps before errors appear is covered in the post worker pool exhaustion. What matters for the alert: the P95 only uses successful checks. If a site tips over so far that requests count as outages, precisely the slowest values drop out of the calculation, and the P95 can fall while the site gets worse. That case is reported by the outage alert, which only confirms an outage when at least two of three locations report it (XICTRAQ manual).
Black Friday: the baseline does not know the calendar
For many shops, peak week brings considerably more visits than an ordinary week. In a representative survey of 1,120 internet users aged 16 and over in Germany in fall 2025, 52 percent planned to use the reduced prices around Black Friday and Cyber Week (Bitkom). Of these, 69 percent planned to look for deals exclusively online (Bitkom). Adding ad budget to the week amplifies the effect; what that budget brings in is covered in the post on ROAS in Black Week.
Official statistics also observe that revenue, and with it load, shifts into November. According to the Federal Statistical Office, part of the Christmas business has increasingly moved forward into November through promotions around Black Friday and Cyber Monday, above all in online and mail-order retail (Destatis). Even before the pandemic, online retail sales rose by 17.6 percent in real terms from October to November 2019 (Destatis).
The baseline sees none of this. It knows hours of week, not a calendar. On Friday, November 27, 2026, it compares the evening hour with the evening hours of November 13 and November 20. If your promotions already start on November 20, part of the pre-load is already in the reference. If the P95 on Black Friday is clearly above what those two Fridays showed, the alert fires. Statistically this is not a false alarm, because the site really is slower than usual. Whether it helps is judged by a brief standard in the SRE book: every page should be actionable (Google SRE Book).
Many shops run their offers across the whole week up to Cyber Monday on November 30. Then every hour of the promotion week compares against a calmer reference, and the number of alerts grows with the length of the promotion. The post on traffic peaks in the fourth quarter recommends measuring response time during the high-load window against your own previous week at the same hour. Automatic mode reaches one step further back and averages over the two previous weeks.
The same basic problem affects every analysis that compares promotion weeks with ordinary weeks, for example A/B tests in Black Week or reports that do not align the year-over-year comparison in the fourth quarter by weekday.
After the peak: two weeks with a raised threshold
More interesting than Black Friday are the Fridays that follow. Because the baseline is recalculated every night over the last 14 days, it consists of exactly two days for every hour of week. On December 4 these are November 20 and 27, on December 11 November 27 and December 4. Only on December 18 has Black Friday dropped out of the window. The same applies to Mondays with Cyber Monday: it sits in the baseline for December 7 and December 14.
The continuation of the example calculation for the Friday evening hour shows how strong the effect is. On normal Fridays the mean is 420 and the standard deviation 80 milliseconds, so the threshold is 660 milliseconds. On Black Friday the values of the same hour average 1,300 milliseconds with a spread of 300 milliseconds. The P95 in the window reaches 1,840 milliseconds, and the alert fires.
Normal Friday, evening hour (example)
μ = 420 ms σ = 80 ms
threshold = max(660 ; 630 ; 500) = 660 ms
Baseline on December 4 and 11: one normal Friday plus Black Friday
μ = (420 + 1,300) / 2 = 860 ms
σ = √((80² + 300²) / 2 + 440²) ≈ 492 ms
threshold = max(2,335 ; 1,290 ; 500) ≈ 2,335 msThe baseline mixes both Fridays in equal parts. The mean rises to 860 milliseconds, and the standard deviation grows to about 492 milliseconds, because the gap between the two Fridays now counts as spread as well. The threshold is then about 2,335 milliseconds, a good three and a half times the usual 660 milliseconds. A real slowdown on December 4, say a P95 of 1,800 milliseconds after a faulty update, would trigger no response-time alert in this example.
In classic process control, control limits are rarely recalculated. Because a control chart compares current performance with past performance, changing the limits frequently would negate any usefulness (NIST/SEMATECH e-Handbook). The adaptive baseline takes the other route and recalculates every night, so that it picks up growth and new load patterns without maintenance. The price: it also picks up special effects such as a promotion week, as long as these are inside the window. Unlike a real incident, though, Black Week can be marked in the calendar in advance.
On the Free plan, raw check data is deleted after 7 days, on the higher plans after 30, 90 or 180 days (XICTRAQ manual). On the Free plan the baseline therefore finds only a single day for most hours of week. On Black Friday it compares with November 20 alone, and afterwards the peak lingers for one week instead of two. In return, the balance of a second day is missing: a single unusual Friday alone determines the threshold for the following week. The plans are compared on the pricing page.
A fixed upper limit for peak week
The simplest countermeasure lies in a rule that applies anyway: a fixed threshold that has been entered takes precedence over automatic mode, and the check then compares the P95 only with that number. The baseline keeps being calculated in the background. As soon as you clear the field again, automatic mode takes over with everything it has learned in the meantime.
The number itself should reflect what your shop demonstrably handles under load and what visitors accept, not the statistics. Reference points are the P95 from a load test or from last year's promotion week and the response time at which orders visibly drop. Without a measurement, start generously and tighten after the first days. How fixed thresholds, severity levels and alert grouping work together is described in the post configuring monitoring alerts. For 2026 this results in a simple schedule:
- By Sunday, November 22: count the values in the window and decide on the fixed upper limit.
- From Monday, November 23: enter the fixed threshold. It applies throughout the promotion week and past Cyber Monday on November 30.
- Through Monday, December 14: leave the fixed threshold in place; the baselines still carry values from the promotion week.
- From Tuesday, December 15: clear the field. The baseline then no longer contains any days from the promotion week.
- Early January: watch for the counterpart. If load drops sharply between Christmas and New Year, the baseline drops too, and the first ordinary working days run against a reference that is too low.
A fixed threshold does not add to automatic mode, it replaces it for as long as it is entered. If you want both, an absolute limit and the detection of unusual deviations, pick the right rule for each week. In peak week the absolute limit counts, in ordinary weeks the deviation.
What the alert tells you
An alert from automatic mode states three values: the P95 in the window, the seasonal normal value for the hour of week and the threshold calculated from it. If an alert in early December shows a normal value of about 860 milliseconds where it usually says 420, the promotion week is still part of the calculation. If the normal value is missing from the alert, the fixed threshold or the 5,000-millisecond fallback was used.
The response-time alert has the severity warning and goes out through the channels you have set up under alerting and notifications. When the P95 falls below the threshold again, a later check closes the alert with an all-clear.
Learn or set: two rules for two questions
Google's SRE book is cautious about learned thresholds. By its own account, the team avoids “magic” systems that try to learn thresholds or automatically detect causality (Google SRE Book). As a principle, it states that the rules that catch real incidents most often should be as simple, predictable and reliable as possible (Google SRE Book).
In that sense the adaptive threshold is not a black box: window, formula and floors are fixed, and every alert based on the baseline states the normal value. What lies inside the 14-day window can be calculated in advance for peak week.
Behind the two rules are different questions. Automatic mode answers whether the site is slower than usual at this hour. The fixed threshold answers whether it is slower than visitors accept. Both questions belong to latency, one of the four golden signals the SRE book names for every user-facing system: latency, traffic, errors and saturation (Google SRE Book). If you frame the second question as a target with an error budget, you will find the approach in the post on error budget and burn rate.
Detects when an hour is clearly slower than the same hour in the two previous weeks, without maintenance.
Measures against a limit you derive from a load test or last year. It takes precedence for as long as it is entered.
Reports failed requests by majority vote of the locations, without a response-time threshold.
Set the peak-week rule before November 23
In ordinary weeks the adaptive threshold remains a good choice, because it knows a website's daily pattern and needs no maintenance. In peak week and the two weeks after, it becomes predictably inaccurate: sensitive against a quiet reference on Black Friday, insensitive against an inflated one afterwards. The fixed upper limit bridges exactly that period.
- Check the window: at least five values, 20 for a calm signal
- Derive a fixed upper limit from a load test or last year and enter it from November 23
- Leave it in place through December 14 and clear it from December 15
- In early December, read the normal value in the alert
- On the Free plan, expect one week of after-effect instead of two
What window, threshold and automatic mode look like in a website's settings is shown in the live demo. The rules in detail are in the uptime monitoring manual.
This post draws on: NIST/SEMATECH e-Handbook of Statistical Methods, sections 6.3.1 “What are Control Charts?”, 6.3.2 “What are Variables Control Charts?” and 1.3.5.11 “Measures of Skewness and Kurtosis”; Google, Site Reliability Engineering, chapter “Monitoring Distributed Systems”; Bitkom, press release “Black Friday: Die Hälfte geht auf Schnäppchenjagd” of November 24, 2025 (survey of 1,120 internet users aged 16 and over); Federal Statistical Office (Destatis), press releases no. N 067 of November 24, 2021 and no. 004 of January 8, 2025; XICTRAQ manual, uptime monitoring chapter. The details on automatic mode reflect the state as of November 2026. All example values are calculated, not measured.