Inflight Documentation

Platform Overview

Inflight is built on a modular architecture designed for reliability, scalability, and extensibility. Each component serves a specific purpose while working together seamlessly.

Architecture

The Inflight platform consists of five core services that work together to provide end-to-end optimization capabilities:

Data Sources

Inflight Agent
+
DynatraceDatadogNew RelicSigNoz

Metrics Collector

Aggregate • Normalize • Store

3-tier

Simulator

Predict • Validate • Backtest

4 modes

AI Advisor

Optimize • Compare • Recommend

Pareto

Output

Dashboard
Alerts
DeploySoon
Monitoring Service runs continuously across all components

How Data Flows

1

Service Registration

The ultra-lightweight Inflight Agent (~5MB) runs as a sidecar in your pods, automatically discovering JVM/Go configurations, resource limits, and registering services with the Metrics Collector. Zero APM dependencies required.

2

Metric Collection & Storage

The Metrics Collector aggregates data from your APM providers into a 3-tier storage system with hot, warm, and cold tiers for optimal performance and cost. Database-driven normalization with intelligent anomaly detection.

3

Multi-Fidelity Simulation

The Simulator uses 4 fidelity modes (Statistical → Hybrid → Full DES → Degraded) to predict configuration impacts. Continuous model calibration with automatic drift detection. Backtester validates predictions against historical data.

4

Multi-Candidate Optimization

The Advisor compares multiple configuration candidates using Pareto frontier analysis. Policy gates enforce tier-based acceptance criteria. Fidelity escalation triggers automatically when confidence is low.

5

Continuous Monitoring

The Monitoring Service tracks your services around the clock, detecting anomalies and correlating alerts to reduce noise and highlight what matters.

Component Details

Learn more about each component and its capabilities:

Inflight Agent

Ultra-lightweight sidecar (~5MB) that registers services and auto-discovers runtime configurations without APM dependencies.

Key Capabilities:

  • Zero APM dependencies—registration only
  • Auto-detect JVM config (GC algorithm, heap, flags)
  • Auto-detect Go config (GOGC, GOMEMLIMIT, GOMAXPROCS)
  • Kubernetes pod-spec discovery for resources
  • Heartbeat-based lease management
  • Multi-APM registration support

Metrics Collector

Multi-APM aggregation hub with 3-tier storage architecture, intelligent normalization, and comprehensive anomaly detection.

Key Capabilities:

  • 3-tier storage architecture (hot/warm/cold)
  • Database-driven normalization rules (runtime configurable)
  • Intelligent anomaly detection
  • Service profile assignments for metric contracts
  • Quality scoring with completeness tracking
  • Real-time metric ingestion

Simulator

Multi-fidelity prediction engine with 4 simulation modes, continuous model calibration, and governance-driven validation.

Key Capabilities:

  • 4 fidelity modes: Statistical → Hybrid → Full DES → Degraded
  • Discrete event simulation engine
  • Model calibration with drift detection
  • Gap detection with automatic interpolation
  • Backtester for rolling-origin model validation
  • Platform-aware safety (Kubernetes, Cloud Run limits)

Advisor

Intelligent orchestration layer with multi-candidate optimization, policy gates, and automatic fidelity escalation.

Key Capabilities:

  • Multi-candidate comparison with Pareto frontier identification
  • Fidelity escalation for low-confidence predictions
  • Policy gates with tier-based acceptance criteria
  • Pluggable optimizer strategy registry
  • Knowledge base with provenance tracking
  • Runtime extractors: JVM, Go, Dependency, Queue

Monitoring Service

Manages the alert lifecycle, from detection to resolution, with intelligent alert correlation.

Key Capabilities:

  • Alert detection and routing
  • Alert correlation
  • Escalation management
  • Resolution tracking
  • Alert analytics

What's Coming

In Development

Deployment Manager

We're building closed-loop deployment capabilities to complete the optimization cycle. The Deployment Manager will automatically implement approved recommendations that meet quality thresholds—taking Inflight from recommendation to action without manual intervention.

  • Automatic deployment of approved changes
  • Quality threshold gates before deployment
  • Continuous feedback loop from production metrics
  • Automatic rollback if metrics degrade

Design Principles

Runtime Agnostic

Designed to work with any runtime environment, not just specific languages or frameworks.

APM Agnostic

Integrates with your existing observability stack rather than requiring a specific provider.

Safety First

Every recommendation is validated through simulation before being presented to you.

Explainable AI

No black boxes. Every suggestion comes with clear reasoning and evidence.