Scaler Configuration

This document describes the comprehensive configuration options for the ICC Scaler service, which implements intelligent autoscaling for Node.js applications using dual reactive and predictive algorithms.

Overview

The ICC Scaler service provides intelligent autoscaling for Node.js applications in Kubernetes environments. It combines a Reactive Autoscaler Algorithm for real-time scaling responses and a Trends Learning Algorithm for predictive scaling based on historical patterns. The system is designed to handle 30-50 scaling events per day with high efficiency and transparency.

Core Architecture

Scaling Algorithms

1. Reactive Autoscaler: Responds in real-time to performance signals (ELU > 90%)
2. Trends Learning Algorithm: Analyzes 30-day history (900-1500 events) for predictive scaling
3. Dual Integration: Combines reactive and proactive scaling decisions

Key Features

• Event Loop Utilization (ELU) and Heap monitoring
• Historical performance clustering
• 3-day half-life decay for recent trend prioritization
• Leader election for multi-instance deployments
• Comprehensive validation and success scoring

Pod Scaling Limits

Caution

These values are very important, the scaler won’t scale up more than PLT_SCALER_MAX_PODS_DEFAULT

Variable	Type	Default	Description
PLT_SCALER_MIN_PODS_DEFAULT	Integer	1	Default minimum number of pods per application
PLT_SCALER_MAX_PODS_DEFAULT	Integer	10	Default maximum number of pods per application

Advanced Scaling Algorithm Parameters

Caution

Changing these configuration parameters will affect A LOT the Scaler algorithms behavior, so please be very careful. In general don’t change the default values in Production unless you know very well what you are doing.

Variable	Type	Default	Description
PLT_SCALER_DEBOUNCE	Integer	10000	Debounce time in milliseconds to prevent rapid scaling
PLT_SCALER_MAX_HISTORY	Integer	10	Maximum number of history events to consider
PLT_SCALER_MAX_CLUSTERS	Integer	5	Maximum number of clusters for historical analysis
PLT_SCALER_ELU_THRESHOLD	Float	0.9	Event Loop Utilization threshold (90%)
PLT_SCALER_HEAP_THRESHOLD	Float	0.85	Heap usage threshold (85%)
PLT_SCALER_POST_EVAL_WINDOW	Integer	300	Post-evaluation window in seconds
PLT_SCALER_COOLDOWN	Integer	300	Cooldown period between scaling operations (seconds)

Execution Control

Variable	Type	Default	Description
PLT_SCALER_PERIODIC_TRIGGER	Integer	60	Periodic trigger interval for metrics-based scaling (seconds)

Algorithm Configuration

Reactive Autoscaler Thresholds

The reactive algorithm uses the following key thresholds:

• ELU Trigger: 90% (configurable via PLT_SCALER_ELU_THRESHOLD)
• ELU Fallback: 95% (ensures critical scaling)
• Heap Trigger: 85% (configurable via PLT_SCALER_HEAP_THRESHOLD)
• Signal Window: 60 seconds of metrics data
• Cooldown Period: 300 seconds between scaling operations

Trends Learning Parameters

The predictive algorithm operates with:

• History Window: 30 days (900-1500 events)
• Half-life Decay: 3 days (λ ≈ 2.674 × 10⁻⁶)
• Execution Frequency: Hourly scheduled runs
• Confidence Threshold: 80% for prediction acceptance
• Prediction Window: 30 seconds for proactive scaling

Validation Metrics

Success scoring combines:

• Responsiveness: 70% weight (ELU/Heap < thresholds post-scaling)
• Resource Optimization: 30% weight (optimal pod utilization)

Scaling Decision Flow

Reactive Scaling Process

1. Signal Detection: Pod subprocess ELU > 90%
2. Metrics Collection: 60 seconds of ELU/Heap data
3. Performance Analysis: Compute pod-level scores using historical clusters
4. Trigger Evaluation: Scale if score > 0.5 or ELU > 95%
5. Prediction Integration: Query trends learning for proactive scaling
6. Scaling Execution: Apply cooldown and pod limits
7. Success Tracking: Monitor post-scaling performance

Predictive Scaling Process

1. Historical Analysis: Process 30-day event history (hourly)
2. Pattern Detection: Identify time-slot probabilities with 3-day decay
3. Sequence Modeling: Model multi-step scaling patterns
4. Validation: Score predictions for responsiveness and optimization
5. Schedule Generation: Create timestamped predictions with confidence scores