MLOps

The MLOps module provides production machine learning operations including A/B testing routers for traffic splitting between model variants, drift detection for monitoring model performance over time, and ground truth integration for measuring real-world accuracy.

Overview

Core Capabilities

• Feature Store -- Define features once and serve them consistently for training and inference, with point-in-time correct training data and low-latency online serving
• A/B Testing Routers -- Split traffic between model variants using weighted random, feature-based, multi-armed bandit, or canary routing strategies
• Drift Detection -- Monitor changes in production data distributions compared to training data baselines using PSI, KS test, Chi-Square, and Jensen-Shannon divergence
• Ground Truth Integration -- Upload actual outcomes to track prediction accuracy over time and detect concept drift
• Performance History -- Track model accuracy trends with configurable granularity (hourly, daily, weekly)
• Alert Configuration -- Set drift thresholds and notification preferences per model

MLOps lives in the sidebar under MLOps, with four pages: Workbench (interactive development), Model Registry, AutoML, and Experiments.

Feature Store

The Feature Store defines the features your models use, keeps their values fresh, and serves them consistently for both training and live inference. It gives you point-in-time correct training data (no label leakage) and low-latency online serving (no training/serving skew). Bind a durable Postgres store for history and, for real-time serving, a Redis store for the latest values. AutoML can train directly on a feature service, and a deployed model can resolve its features automatically at inference.

Learn more about the Feature Store -->

A/B Testing Routers

Compare model variants in production with traffic splitting across four routing strategies:

Routing Strategies

Strategy	Description
Weighted Random	Split traffic by percentage between variants. A variant with weight 70 gets 70% of traffic relative to total enabled variant weight.
Feature-Based	Route requests based on input feature values using priority-ordered rules with operators (equals, not_equals, in, not_in, greater_than, less_than, contains, regex).
Multi-Armed Bandit	Automatically balance exploration and exploitation to route more traffic to better-performing variants. Supports epsilon-greedy, Thompson sampling, and UCB1 algorithms.
Canary	Gradually roll out a new model variant by routing a configurable percentage of traffic to the canary while the rest goes to the control variant.

Router Lifecycle

Routers follow this lifecycle:

registered --> deploying --> running --> stopped
                              |
                              v
                            failed

Status	Description
registered	Router created but not yet deployed
deploying	Deployment to Kubernetes is in progress
running	Router is live and serving predictions
stopped	Router is paused, preserving configuration
failed	Deployment failed with errors

Access Control

• Owner -- The user who created the router (full modify/delete access)
• Public -- Router can be marked as public for all users to view
• Shared -- Router can be shared with specific users by ID
• Admins -- Administrators have full access to all routers

Learn more about A/B Testing -->

Drift Detection

Monitor model performance and detect distribution changes in production data:

Statistical Tests

Metric	Use Case	Interpretation
PSI (Population Stability Index)	Overall distribution shift	< 0.1: OK, 0.1-0.2: Warning, > 0.2: Alert
Kolmogorov-Smirnov Test	Numerical feature drift	p-value < 0.05 indicates significant drift
Chi-Square Test	Categorical feature drift	p-value < 0.05 indicates significant drift
Jensen-Shannon Divergence	General distribution comparison	0-1 scale, higher values indicate more divergence

Drift Detection Workflow

1. Create a reference baseline from your training data feature distributions
2. Log predictions as your model serves requests in production
3. Run drift analysis to compare current production data against the baseline
4. Review results showing overall drift score and per-feature metrics
5. Upload ground truth to measure actual prediction accuracy over time

Drift Status Levels

Status	PSI Range	Action
ok	PSI < 0.1	No action needed. Data distribution is stable.
warning	0.1 <= PSI < 0.2	Investigate potential causes of drift.
alert	PSI >= 0.2	Significant drift detected. Consider retraining the model.

Learn more about Drift Detection -->

Router Methods

Method	Description
routers.create	Create a new router with strategy and variants
routers.get	Get a router by ID (access-controlled)
routers.list	List accessible routers with optional filters
routers.update	Update router name, description, or tags
routers.delete	Soft-delete a router (stops deployment first)
routers.deploy	Deploy router to Kubernetes via AI Gateway
routers.stop	Stop a running router
routers.start	Start a stopped router (re-deploys)
routers.updateVariantWeight	Change a variant's traffic weight (0-100)
routers.toggleVariant	Enable or disable a variant (minimum 2 must remain enabled)
routers.getMetrics	Get aggregated metrics with per-variant breakdown
routers.getPredictions	Get prediction history with pagination
routers.recordFeedback	Record reward feedback for bandit optimization
routers.updatePermissions	Update public/shared access settings

Drift Detection Methods

Method	Description
drift.logPrediction	Log a prediction with features and output for drift tracking
drift.getPredictions	Get prediction logs with date and ground truth filters
drift.addGroundTruth	Upload a single ground truth record
drift.uploadGroundTruth	Bulk upload ground truth from CSV/JSON
drift.getUnmatchedPredictions	Find predictions without ground truth
drift.createBaseline	Create a reference baseline from feature data
drift.getActiveBaseline	Get the active baseline for a model
drift.setActiveBaseline	Set which baseline is active
drift.getBaselines	List all baselines for a model
drift.analyze	Run drift analysis comparing production data to baseline
drift.getLatestResult	Get the most recent drift analysis result
drift.getResults	Get drift result history with optional status filter
drift.getAlertConfig	Get alert configuration for a model
drift.updateAlertConfig	Update alert thresholds and notification settings
drift.getPerformanceHistory	Get accuracy history with configurable granularity

Getting Started

Setting Up A/B Testing

1. Register your model variants in the Model Registry
2. Create a router with at least 2 variants and a routing strategy
3. Deploy the router to get a prediction endpoint
4. Route your application's requests through the router endpoint
5. Monitor metrics and adjust variant weights based on results

Setting Up Drift Detection

1. Export feature distributions from your training dataset
2. Create a reference baseline using drift.createBaseline
3. Configure prediction logging so production predictions are tracked
4. Run periodic drift analysis using drift.analyze
5. Optionally upload ground truth data to track accuracy over time
6. Configure alert thresholds to be notified of significant drift

Best Practices

A/B Testing

• Start with small traffic percentages (5-10%) for new variants before increasing
• Use multi-armed bandit routing when you want to minimize exposure to underperforming variants
• Define success metrics before starting a test
• Ensure statistically significant sample sizes before drawing conclusions
• Use canary routing for cautious rollouts of new model versions

Drift Detection

• Create baselines immediately after training while training data is available
• Run drift analysis at regular intervals (daily or weekly) to catch issues early
• Collect ground truth data wherever possible to measure real accuracy, not just distribution shifts
• Investigate warning-level drift before it becomes critical
• When significant drift is confirmed, retrain the model and create a new baseline
• Use A/B testing routers to safely compare retrained models against current production versions