Data Analytics – How It Works¶
The Idea¶
Data Anomalies asks whether today deviates from expectation. Data Analytics asks what the last N observations, taken together, are doing — and answers it with a straight line fitted through them.
Two properties of that line matter:
- Its slope is the trend — where the metric is heading.
- The scatter around it is the volatility — how consistently it gets there.
Both are computed from statistics profiling has already collected, so the module adds analytical depth without a second pass over your data.
Defining an Analytics Rule¶
A rule names four things:
| Part | Meaning |
|---|---|
| The series | Which table, column, and statistic to analyse |
| The metric | Absolute change, relative change, absolute volatility, or relative volatility |
| The window | A range and a unit — how much history the line is fitted through |
| The limits | Optional lower and upper bounds on the resulting value |
Window units¶
| Unit | Meaning |
|---|---|
| DAY | The last N days |
| WEEK | The last N weeks |
| MONTH | The last N months |
| OBSERVATION | The last N observations, whenever they happened |
OBSERVATION is the one to use for irregular series. A monthly delivery analysed over "the last 30 days" has one or two points; analysed over "the last 12 observations" it has a year of behaviour. Calendar units are the better choice when the rate matters — a source whose delivery frequency itself is changing tells you more per day than per observation.
The Four Metrics¶
A regression is fitted with time on the x-axis, scaled so the window spans the unit interval. That scaling is what makes the metrics comparable: a slope is expressed per window, not per second, so the same limit means the same thing on a 30-day and a 90-day rule.
| Metric | Definition | Reads as |
|---|---|---|
| Absolute Change | The slope of the fitted line | How much the metric moves across the window, in its own units |
| Relative Change | Slope ÷ intercept, as a percentage | The same movement as a percentage of where the window started |
| Absolute Volatility | Mean absolute deviation from the fitted line | Typical distance between observation and trend, in the metric's own units |
| Relative Volatility | That deviation ÷ the mean of the series, as a percentage | Scatter as a percentage of the typical level |
Choosing absolute or relative¶
Relative metrics are what you want in almost every case where series are compared or one rule is applied across many columns — a 5 % drift means the same thing on a table of 500 rows and one of 5 million. Absolute metrics are right when the quantity itself has meaning: "this null count must not grow by more than 100 per month" is an absolute statement, and expressing it as a percentage would be an awkward way to say it.
Relative metrics are guarded against a near-zero denominator: when the intercept, or the mean of the series, is effectively zero, the relative value is reported as zero rather than exploding. A metric that spends time near zero should be monitored with its absolute counterpart.
Evaluation¶
Each computed value is compared against the rule's limits:
| Condition | Status |
|---|---|
| Below the lower limit, or above the upper limit | Failed |
| Otherwise | Passed |
A limit left empty is simply not evaluated, so a rule with neither limit set records the value and never fails. That is a legitimate configuration — it gives you the trend series in the interface without alerting on it, which is a good way to observe a metric's normal range before deciding what the limit should be.
There is no Uncertain status here. Unlike anomaly detection, whose band is learned and therefore has a natural inner and outer edge, an analytics limit is a line a person drew.
Statuses roll up the same way as everywhere else — check → attribute → dataset → data source, worst result winning.
Practical Setup¶
- Create the rule with no limits. Let it run for a few weeks and watch what the metric actually does on your data.
- Set the limit outside the observed range, not at the edge of it. A limit set to the highest value you have seen will fire on the next ordinary week.
- Prefer one-sided limits where only one direction is a problem. A rising null rate is a defect; a falling one is a fix, and does not need an alert.
- Use volatility for pipeline health, trend for data health. Rising volatility on a stable metric is usually an infrastructure story — retries, contention, an unreliable upstream job. A moving trend is usually a data or business story.
Read the two together
Flat trend with rising volatility means instability, not change. A moving trend with low volatility is an orderly change — growth, a migration, a decommissioning. A moving trend and rising volatility is usually a defect: partial loads, duplicated batches, a source in transition.
Related Pages¶
- Data Analytics – Introduction
- Data Analytics – Use Cases
- Profiling – The Foundation
- Statistics
- Statuses and Alerts