Thresholds and baselines
Every check in Excubius starts from a baseline: a published, named value such as Microsoft's guidance or your own recovery objectives. Entities and single resources can override that value, and the screens always show where the value in force comes from. This article explains baselines and their sources, how the effective value is resolved, how severity and the for duration decide when an alert opens, and how checks are switched on and off.
How the effective threshold is resolved
Excubius walks from the catalog baseline down the entity tree to the resource. Each level can replace a value; the most specific one wins. Values are resolved key by key: in the example, the warning level comes from Production while the critical level comes from the resource itself.
The keys a level can override are:
| Key | Means |
|---|---|
warn | The warning rule |
crit | The critical rule |
expect | The expected state, for checks that compare states |
window, sigma | The learning window and band of an adaptive check |
A level can also turn a rule off. An entity that turns off the warning level of a check stops warning alerts for that check on every resource below it, until a lower level sets the rule again.
Baselines and their sources
Every check in the catalog names the source of its baseline, with a reference and, often, a note that explains the value. The Checks tab of a resource shows it under Baseline: when you edit a check, and the alert detail shows the note, so nobody has to guess why a limit is where it is.
| Source | Typical baselines |
|---|---|
| Microsoft documentation | SQL Server CPU 80% and 90%, disk latency of 20 ms for reads and 10 ms for log writes, availability group health states |
| Azure Monitor recommended alerts | Virtual machine CPU above 80%, available memory below 10%, Azure SQL DTU above 80%, any deadlock |
| CIS Benchmark | sa disabled, surface-area features off, privileged group membership, audit configured |
| Community practice | Page life expectancy, blocking, log space, queue sizes |
| Your recovery objectives (RPO / RTO) | Age of the last full and log backup, estimated data loss of an availability group, Data Factory daily deadline |
| Vendor documentation | PostgreSQL transaction ID wraparound, Linux errata, Azure services |
A baseline is a starting point, not a verdict. Express your recovery objectives as entity defaults of the backup and data-loss checks, for example on Ciguapa Inc., and every resource below that entity follows them.
Who changes what
| Where | Who | How |
|---|---|---|
| One resource | Editor or higher on its entity | The Checks tab of the resource: thresholds for this resource only. A level you leave alone keeps its inherited value. Back to inherited thresholds removes the resource's own values |
| An entity and everything below it | Entity admin on that entity | Settings › Entities, threshold defaults: for each level, Inherit, Set here or Turn off here |
Every change is recorded in the audit log with who made it and when. Entity defaults are available for checks that compare a value with limits; checks that compare states keep the rules of the catalog, and the editor says so.
Severity and the for duration
A check can carry two rules, warn and crit. Each rule has an operator (>, >=, <, <=, ==, !=, in, not_in or matches), a value and an optional for duration such as 30s, 15m, 1h or 1d.
- A condition must hold for its whole
forduration before the alert opens. A short spike that clears earlier never becomes an alert. - The alert takes the worst level whose condition has held long enough. When a warning becomes critical, the same alert escalates; it does not open a second one.
- Recovery is immediate: as soon as the condition no longer holds, the alert resolves.
Not every check compares a number with limits. The catalog uses a few kinds of evaluation:
| Kind | How it decides |
|---|---|
| Threshold | The rules on the measured value, for example CPU percentage |
| Age | The rules on an age or on the days left before something expires. Something that never happened, such as a database never backed up, counts as infinitely old |
| Count | The rules on a number of rows, for example blocked sessions. No rows is a count of zero, so the alert resolves |
| State | Each field of the result must match the expected state, for example a replica that must be HEALTHY |
| Change | A counter becomes a rate per second or per minute and the rules apply to the rate. For other values, any change from the previous run, such as a failover or a new member of sysadmin, is an event that stays open until someone resolves it |
| Adaptive | Learns the normal range of the value; see below |
Probes that exercise a real path, such as a linked server or a web endpoint, turn critical after a number of failed attempts in a row: one by default, three for linked servers.
Adaptive baseline
Some throughput metrics have no sensible fixed limit: batch requests per second, the duration of a Data Factory pipeline, application throughput. Their checks use an adaptive baseline, which compares the value with the resource's own history over a window, for example 14 days with a band of two standard deviations.
In this release adaptive checks only learn. Their values are stored, their state shows Learning, and they raise no alert yet. They never change the fixed thresholds of other checks.
Switching checks on and off
When a resource is added, Excubius gives it every check of its type that the catalog marks as on by default, for the modules your installation has switched on. A few catalog checks are off by default and are added to a resource on request.
- An Editor switches a check off or on for one resource in the Checks tab, for example on a lab server. A switched-off check stays in the list, greyed, so nobody forgets it exists.
- An editor can also change how often a check runs on that resource, from 10 seconds to 7 days. Left empty, the check runs on the catalog's interval, between 30 seconds and 24 hours.
- To stop a check alerting across a whole branch while keeping its history, an entity admin turns its levels off on the entity.
Your choices survive rediscovery and upgrades: Excubius only adds missing default checks and never touches existing ones.