> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dqlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Alert Thresholds

> How to configure Auto, Limit, and Custom alert thresholds for Prizm metrics — including multi-criteria custom rules, sensitivity tuning, and per-condition actions.

<script type="application/ld+json">
  {`{
            "@context": "https://schema.org",
            "@type": "TechArticle",
            "headline": "Alert Thresholds",
            "description": "How to configure Auto, Limit, and Custom alert thresholds for Prizm metrics - including multi-criteria custom rules, sensitivity tuning, and per-condition actions.",
            "url": "https://docs.dqlabs.ai/architecture/alerts/thresholds",
            "publisher": {
              "@type": "Organization",
              "name": "DQLabs Inc",
              "logo": "https://media.brand.dev/332adc35-5bc4-4d2b-bf78-256aa4a5e414.svg"
            }
            }`}
</script>

A threshold defines when a metric value is considered anomalous and should fire an alert. Prizm provides three threshold types — **Auto**, **Limit**, and **Custom** — selectable from the metric's Configuration tab.

## Threshold types

| Type       | Best for                                                                                    | How it works                                                                         |
| ---------- | ------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ |
| **Auto**   | Metrics with variable historical patterns                                                   | ML-based anomaly detection using a statistical baseline. Learns from run history.    |
| **Limit**  | Metrics with known, fixed acceptable ranges                                                 | A single explicit bound defined by the user. Fires when the value crosses the limit. |
| **Custom** | Metrics needing multiple named alert criteria with different rules, priorities, and actions | One or more named criteria evaluated independently per run.                          |

***

## Auto threshold

Auto thresholds use statistical models trained on a metric's historical run values to detect anomalies. Prizm calculates a rolling mean and standard deviation, then flags values that deviate beyond the configured sigma bands.

| Band              | Default action      |
| ----------------- | ------------------- |
| Within 2σ         | No alert            |
| Between 2σ and 3σ | Medium / High alert |
| Beyond 3σ         | Critical alert      |

### Sensitivity and lookback

| Setting                         | Effect                                       |
| ------------------------------- | -------------------------------------------- |
| **Tight** (lower σ multiplier)  | Narrower bands — more sensitive, more alerts |
| **Standard**                    | Default bands                                |
| **Loose** (higher σ multiplier) | Wider bands — less sensitive, fewer alerts   |

The **lookback window** (7 days to 1 year) controls what historical data the baseline mean and standard deviation are computed from. A shorter window makes the baseline more reactive to recent changes; a longer window gives a more stable baseline.

Each sigma band has its own **action** — for example, Medium band can notify the owner while Critical blocks the pipeline:

| Action             | Effect                                                     |
| ------------------ | ---------------------------------------------------------- |
| **No action**      | Alert recorded, no notification triggered                  |
| **Notify owner**   | Sends a notification to the asset owner and alert audience |
| **Create issue**   | Automatically creates a Prizm issue linked to this alert   |
| **Block pipeline** | Signals downstream orchestrators to halt execution         |

<Note>
  Auto thresholds require a minimum number of historical runs before they become active. During this period the metric is in **Learning** mode and no alert fires.
</Note>

Prizm also supports **drift detection** — if the metric value shifts gradually over time rather than spiking, drift mode catches the sustained trend before it becomes a hard threshold breach.

***

## Limit threshold

Limit thresholds let you define an explicit bound based on known business rules — useful when historical data is not a reliable guide for what "normal" looks like.

### Condition types

| Condition                    | Example                                     |
| ---------------------------- | ------------------------------------------- |
| **Equal to**                 | `= 0` (zero-tolerance null check)           |
| **Not equal to**             | `≠ 100` (completeness must be exactly 100%) |
| **Greater than**             | `> 500`                                     |
| **Greater than or equal to** | `>= 1000`                                   |
| **Less than**                | `< 0.95` (95% completeness SLA)             |
| **Less than or equal to**    | `<= 0`                                      |
| **Between**                  | Between `0.98` and `1.02`                   |

### How to set a Limit threshold

<Steps>
  <Step title="Open metric configuration">
    Navigate to the metric and open the **Configuration** tab.
  </Step>

  <Step title="Select Limit">
    Under **Threshold**, select **Limit** and choose a condition type.
  </Step>

  <Step title="Set the value">
    Enter the bound. For **Between**, set both lower and upper bounds.
  </Step>

  <Step title="Set alert severity">
    Choose the severity that fires when the limit is breached: Critical, High, Medium, or Low.

    <Frame>
      <img src="https://mintcdn.com/dqlabs/52LiJLU_AuJjpUWw/images/image-132.png?fit=max&auto=format&n=52LiJLU_AuJjpUWw&q=85&s=952d374939be6db27b085f9f42c90c60" alt="Image" width="2530" height="932" data-path="images/image-132.png" />
    </Frame>
  </Step>

  <Step title="Save">
    Click **Save Changes**. The threshold takes effect on the next metric run.
  </Step>
</Steps>

<Note>
  **Mark as Normal** is not available for Limit threshold metrics. Alerts from Limit thresholds can only be Active or Resolved.
</Note>

***

## Custom threshold

Custom threshold lets you define one or more named **criteria**, each with its own rule type, conditions, priority, and action. Criteria are evaluated independently — if multiple criteria are satisfied in a single run, Prizm generates an alert for each one.

Custom threshold is available for **Query metrics** and **Conditional metrics**.

### Criteria structure

Each criterion requires:

| Field             | Description                                                      |
| ----------------- | ---------------------------------------------------------------- |
| **Criteria name** | A label for this alert condition (e.g., `High Sales Alert`)      |
| **Priority**      | Severity if this criterion fires: Critical, High, Medium, or Low |
| **Rule type**     | The type of check to perform                                     |
| **Condition**     | The comparison logic                                             |
| **Action**        | What Prizm does when this criterion fires                        |

### Rule types

**Row Check** — Evaluates a condition against individual rows. Fires if any row matches. Example: `QTY_SOLD Between 10 and 20`

**Column vs Column** — Compares two columns row by row. Fires if any row satisfies the comparison. Example: `ACTUAL_REVENUE < FORECAST_REVENUE`

**Aggregate** — Applies an aggregate function and evaluates the result against a threshold. Example: `SUM(REVENUE) < 1000`

### Condition operators

| Operator                                 | Description                                           |
| ---------------------------------------- | ----------------------------------------------------- |
| **Greater / Less / Equal to**            | Standard comparison against a fixed value             |
| **Between**                              | Value falls within a min–max range                    |
| **Changed by more than**                 | Value changed by more than N units since the last run |
| **Changed by more than % from last run** | Value changed by more than N% since the last run      |

### Value unit toggle

For numeric conditions, a three-way toggle controls how the value is interpreted:

| Symbol | Meaning                                          |
| ------ | ------------------------------------------------ |
| `#`    | Absolute count                                   |
| `%`    | Percentage of total rows or metric value         |
| `Δ%`   | Percentage change from the previous run's result |

### Multiple criteria example

| Criterion         | Rule                                     | Priority | Action       |
| ----------------- | ---------------------------------------- | -------- | ------------ |
| High Sales Alert  | Row Check — `QTY_SOLD Between 10 and 20` | High     | Notify owner |
| Low Revenue Alert | Aggregate — `SUM(REVENUE) < 1000`        | Medium   | Create issue |

If both conditions are met in the same run, both alerts fire independently.

### Advanced options

**Apply to filtered rows only** — When conditional rules filter the metric (e.g., `WHERE STATUS = 'closed'`), enabling this option scopes the threshold evaluation to only the filtered rows rather than the full table.

**Consecutive runs requirement** — Requires the condition to be met for N consecutive runs before an alert fires. Prevents single-run spikes from triggering alerts while still catching persistent issues.

***

## Switching between threshold types

You can switch a metric between Auto, Limit, and Custom at any time. When switching away from Auto, the historical learning data is preserved — switching back resumes from the last model state.

## Thresholds and quality scoring

Threshold breach status feeds into the asset quality score. A metric in breach reduces the asset's score by its configured **Weightage %**. The score contribution is restored when the metric returns to a valid range or the alert is manually resolved.