The training process in Blindata is dynamic and adaptive, tailored to ensure optimal anomaly detection for time series data. The system initiates the training of temporal series under various conditions: first, upon the insertion of a new record; second, when a change in trend or anomaly is detected; and third, either periodically within a predefined time window or dynamically, adjusting to the statistical properties of the series.

The model for a given metric can exist in different states, reflecting the training status.

• Inactive: When model training is disabled, or it has not started the state is set to INACTIVE.
• Training: When training data is insufficient, indicating that the forecast is not yet available, the state is TRAINING.
• Trained: Once the model has been successfully trained, and the forecast is ready for use, the state transitions to TRAINED.
• Error: In case an error occurs during the training process, rendering the forecast unavailable, the state is marked as ERROR.
• Running: For ongoing training processes, the state is set to RUNNING, indicating that the system is actively updating the model based on the latest data.

This multifaceted approach to training ensures that Blindata remains agile and responsive to changes in data patterns, optimizing its anomaly detection capabilities.