Fix error

Author: fmind

docs/7. Observability/7.2. Alerting.md

@@ -9,6 +9,7 @@ description: Understand the critical role of alerting in AI/ML monitoring, learn
 [AI/ML Alerting](https://www.datadoghq.com/solutions/machine-learning/) is the practice of automatically notifying stakeholders when a production machine learning model's performance or behavior deviates from established norms. It functions as an early warning system, transforming monitoring data into actionable notifications.
 
 An effective alerting strategy is built on three pillars:
+
 - **Defining Triggers**: Establishing precise conditions that signal a potential issue, such as a sudden drop in accuracy or a significant shift in input data.
 - **Routing Notifications**: Ensuring the right individuals or teams are notified based on the alert's nature and severity.
 - **Choosing Channels**: Selecting the most effective communication tools (e.g., Slack, email, PagerDuty) to deliver the alert.
@@ -18,6 +19,7 @@ An effective alerting strategy is built on three pillars:
 Alerting is non-negotiable for maintaining the reliability and performance of production AI/ML models. It moves teams from a reactive to a proactive stance on model maintenance.
 
 Key benefits include:
+
 1.  **Immediate Issue Detection**: Alerts drastically reduce the mean time to detection (MTTD), allowing teams to address problems before they impact users or business outcomes.
 2.  **Proactive Maintenance**: By catching issues like model drift or performance degradation early, alerts trigger necessary interventions like model retraining or system adjustments, preventing larger failures.
 3.  **Data-Driven Decisions**: Alerts provide concrete evidence to justify actions such as model rollbacks, hyperparameter tuning, or infrastructure scaling.
@@ -28,6 +30,7 @@ Key benefits include:
 Alert triggers must be carefully selected to be meaningful and actionable. Overly sensitive triggers lead to alert fatigue, while insensitive ones defeat the purpose of monitoring.
 
 Common and effective alert conditions include:
+
 - **Performance Degradation**: A statistically significant drop in a key evaluation metric (e.g., F1-score, MAE, AUC) below a predefined threshold.
 - **Data and Concept Drift**: A significant statistical divergence (e.g., detected by a Kolmogorov-Smirnov test) between the production data distribution and the training data distribution.
 - **Prediction Anomalies**: The model generates a high rate of outlier predictions or fills a particular prediction class with unusually high or low frequency.
@@ -39,6 +42,7 @@ Common and effective alert conditions include:
 Setting the right threshold is a balance between sensitivity and practicality. A threshold that is too tight will generate constant noise, while one that is too loose may miss critical incidents.
 
 Consider these approaches:
+
 - **Static Thresholds**: A fixed value based on business requirements or historical performance (e.g., "alert if accuracy drops below 90%"). This is simple to implement but can be rigid.
 - **Dynamic Thresholds**: Thresholds that adapt based on historical patterns, such as a moving average or seasonality (e.g., "alert if prediction latency is 3 standard deviations above the weekly average"). This method is more resilient to normal fluctuations.
 - **Canary-Based Thresholds**: When deploying a new model version, alert if its performance is significantly worse than the currently stable production version.