Model Monitoring Dashboard

The Monitoring Dashboard provides a unified interface for tracking your production model’s health, consolidating performance metrics and drift analysis into a single view.

Model Monitoring Dashboard
Figure 1. Model Monitoring Dashboard Interface

Header

The header provides immediate context regarding the data being analyzed

Dashboard Header
Figure 2. Dashboard Header and Metadata

Monitored Pipeline

The active pipeline deployed in production. It contains the specific version of the model currently being observed for performance stability and data integrity.

Dataframe

The production dataset containing live inference data. This dataframe includes the features used during the training phase and the resulting model predictions.

Last Execution

Displays the timestamp of the most recent data execution.

From/To timestamp Filters

Use the date picker (e.g., 1 Jan, 2015 to 20 Jan, 2026) to isolate specific time ranges for retrospective analysis.

Performance Summary KPIs

Provides health indicators.

KPI Overview
Figure 3. Performance Health Overview

#CHUNKS

Total number of chunks in monitored data frame.

HEALTHY

Total number of chunks in monitoring period with zero alerts for all selected metrics.

TOTAL ALERTS

Total alerts numbers.

ALERTS PER METRIC

Displays the number of unique metrics (e.g., MAE, RMSE) that have currently triggered at least one active alert.

AFFECTED CHUNKS

A percentage-based visual representation of overall number of chunks with at least one alert.

Performance Alerts

The Performance Alerts section provides a detailed, tabular audit of model health across individual data chunks.

Performance Alerts Table
Figure 4. Performance Alerts Table

Table Controls

Before analyzing the data, use the interface controls to filter the results:

  • Show healthy chunks: Toggle this switch to filter out stable data and focus exclusively on chunks that have triggered alerts.

  • Metric Filter: Filter the table to display specific performance indicators such as MAE, RMSE, and etc.

  • Realized and Estimated: Toggle between realized performance, estimated performance or both.

  • Sort By: Reorder the data by Oldest or Latest.

Data Column Definitions

CHUNK KEY

The unique identifier for the data (e.g., 2017-02-16 11:00).

CHUNK START/END INDEX

The specific row indices within your dataset of the chunk.

CHUNK START/END

The timestamps defining the beginning and end of the chunk’s monitoring period.

METRIC

The selected performance metric (e.g. msle).

VALUE

The calculated numerical result for the metric within this specific chunk.

SOURCE

Indicates the methodology used to calculate each selected metric’s value: Realized (derived from ground truth labels) or Estimated (derived from estimation algorithms).

CONFIDENCE BAND

The statistically expected range for the metric. Values outside this band typically trigger alerts.

FINAL ALERT

A boolean indicator (true/false) confirming if the chunk officially exceeded its critical threshold.

Performance Through Time

The Performance Through Time visualization provides a temporal analysis of your model’s performance.

Performance Through Time Chart
Figure 5. Performance Through Time Chart
Data Selection Toggle

In the top-right corner, use the dropdown menu to filter the data source for the chart:

  • Realized and Estimated: Displays both the calculations and the monitoring model’s estimations.

  • Estimator: Displays predicted model performance based on analysis of input data when ground truth is unavailable.

  • Calculator: Displays realized model performance derived from a direct comparison between model predictions and ground truth labels.

Chart Legend & Elements

  • ● Estimated: Represented by the purple line, this metric displays the predicted performance level based on analysis of input when ground truth is unavailable.

  • ● Realized: Represented by the blue line, this metric displays the actual performance calculated by comparing model predictions against confirmed ground truth labels.

  • - - Confidence Band: The dashed boundary representing the Upper and Lower Threshold.

  • ◆ Alert: Represented by the red diamond, this identifies specific intervals where the model’s performance has breached a defined threshold or confidence bound.

Threshold Boundaries

The horizontal red dashed lines indicate your Critical Limits.

Any data point that crosses these boundaries (e.g., the 1072.3 upper limit shown in the image) are automatically flagged in the Performance Alerts table.

Univariate Data Drift

Monitors the stability of individual features independently.

Univariate Data Drift Table
Figure 6. Univariate Data Drift Table

Drift Audit Controls

Use these filters to refine your data drift analysis:

  • Method Toggle: Switch between Continuous Method and Categorical Method to view the appropriate statistical tests.

  • Metric Selection: Choose the specific algorithm, such as Jensen Shannon, to visualize the magnitude of the shift.

  • Sort By: Reorder the analysis by Oldest or Latest.

Data Column Definitions

CHUNK INDEX

The sequential number assigned to each data window processed by the system.

CHUNK KEY

The unique identifier for the data (e.g., 2017-02-16 11:00).

CHUNK START/END INDEX

The specific row indices within your dataset of the chunk.

CHUNK START/END DATE

The calendar dates defining the beginning and ending of the monitored model.

FEATURE COLUMNS

Individual columns for each monitored feature (e.g., PRICE NEW, CAR AGE). Cells are color-coded based on the calculated drift score.

  • Green Cells indicate that the individual columns (e.g., PRICE NEW, CAR AGE) is stable and matches.

  • Red/Amber Cells indicate that the individual columns (e.g., PRICE NEW, CAR AGE) have a significant statistical shift.

Data Drift Through Time

Data Drift Through Time Chart
Figure 7. Data Drift Temporal Visualization

Analysis Toggles

Customize the chart view to focus on specific drift dimensions:

  • Univariate vs. Multivariate

    • Univariate: Monitors individual features independently to detect changes in their specific distributions over time.

    • Multivariate: Analyzes the collective behavior and relationships between all selected features simultaneously.

  • Method & Metric Selection:

  • Feature Selection: Use the column dropdown to select which specific features to display on the chart trend line.

Chart Legend & Indicators

  • ● Feature A (e.g., price_new): Represented by the light blue trend line.

  • ● Feature B (e.g., car_age): Represented by the purple trend line.

  • ◆ Alerts:

    • Univariate Alert: Indicates that a specific feature’s distribution has statistically shifted beyond the predefined threshold.

    • Multivariate Alert:  Indicates a detected shift in the joint distribution of all analyzed features.

Prev
Configure Project