A.1. How to read the results
Criteria for confidence bands, correlation coefficients and anomaly verdicts
This collects the criteria for deciding how far to trust a number when one comes back.
1. The confidence band — forecasting
A forecast chart draws one line with a band above and below it.
The width of the band is the degree of confidence.
| Shape of the band | How to read it |
|---|---|
| Narrow | The trend was steady — worth relying on |
| Wide | It was erratic — treat it as indicative only |
| Widens with distance | Normal. The further ahead, the less certain |
When reading off when the limit is reached
Look at when the top of the band reaches it, not the forecast line. You have to take the worst case to be prepared for it.
2. The correlation coefficient — correlation analysis
How far two metrics moved together, as a value between -1 and 1.
| Value | How to read it |
|---|---|
| 0.7 and above | They clearly moved together |
| 0.4 to 0.7 | They moved together to some extent |
| Below 0.4 | Almost no relationship |
| Negative | They moved in opposite directions |
A negative value is not meaningless. Some pairs are supposed to move in opposite directions, such as "free memory ↓ response time ↑". Judge the strength by the absolute value, not the sign.
A high value does not make it the cause
Both may have been driven by a third factor. A correlation result is a narrowed list of candidate causes, not a confirmed cause (Chapter 402).
3. Lead and lag — which moved first
An A → B combination also tells you which one moved first.
- If A moved first, look into A first
- If they moved at the same time, both may be consequences of something else
4. The anomaly verdict — anomaly detection
It calculates what "normal" is and judges by how far a value departs from it.
| Indication | Meaning |
|---|---|
| A large departure | Clearly different from normal |
| Sitting on the boundary | Ambiguous — check what was happening at that time |
What to think of first
Deployments, batches and events are flagged even though they are normal. Look more closely only when nothing comes to mind (Chapter 404).
5. Execution plan cost — SQL
cost is an estimate made by the database.
- It is not seconds
- It cannot be compared with the cost of another query or another database
- The
■bar is a size relative to that plan
It is a value for seeing which part of the same plan is heavier (Chapter 503).
What to suspect first
When a result looks wrong, work through it in this order.
The habit of not taking an answer at face value
- Verify numbers and any command it tells you to run
- Expand
CogentAI's thoughtsabove the answer - Look at the charts and tables it based the answer on
Next
- How it is calculated — Appendix: How forecasts are calculated · Appendix: How anomalies are judged · Appendix: How correlation is measured
- How far it can be trusted — Appendix: Can you trust it