sentinel

Sentinel

Predictive failure detection and autonomous orchestration for Kubernetes edge nodes.

What it does

Sentinel runs on each edge node and does three things:

1. Predicts failures - Monitors CPU thermals, memory pressure, disk I/O, and network health. Uses lightweight statistical models to predict failures before they happen.

2. Survives partitions - When nodes lose contact with the control plane, they form a local consensus group and continue operating autonomously.

3. Takes action - Triggers preemptive workload migrations, cordons unhealthy nodes, and logs decisions for reconciliation when connectivity returns.

Why

Kubernetes assumes the control plane is always reachable. Edge environments break this assumption constantly - spotty networks, power issues, harsh conditions. When a node goes "Unknown", Kubernetes stops making decisions for it.

Cloud AIOps tools assume unlimited resources. Edge nodes run K8s + workloads + monitoring in 4GB RAM with no datacenter cooling.

Sentinel fills the gap.

What's different

Most edge/IoT platforms focus on deployment and connectivity. Most AIOps tools assume cloud-scale resources. Sentinel combines ideas from both:

• Predictive, not reactive - Catches thermal throttling and memory pressure before they cause failures
• Autonomous, not dependent - Continues operating during control plane partitions instead of going dark
• Lightweight, not bloated - Runs statistical models in <64MB RAM, no GPUs or cloud inference needed
• Consensus-aware - Nodes coordinate decisions locally using Raft-lite when disconnected

Quick Start

# Build
make build

# Run locally
./bin/predictor --node=my-node

# With config file
./bin/predictor --config=config.yaml

# Deploy to cluster
helm install sentinel ./deploy/helm/sentinel

Configuration

Sentinel supports YAML/JSON config files with CLI flag overrides:

node:
  name: edge-node-1

server:
  listen: ":9101"
  metrics: ":9100"

predictor:
  interval: 1s
  warn_threshold: 0.3
  critical_threshold: 0.7
  risk_weights:
    thermal: 0.3
    memory: 0.3
    disk: 0.2
    network: 0.2

consensus:
  enabled: true
  addr: ":9200"
  peers:
    - edge-node-2:9200
    - edge-node-3:9200

logging:
  level: info
  format: json

CLI flags:

--config          Config file path
--node            Node name
--listen          API listen address (default :9101)
--metrics         Prometheus metrics address (default :9100)
--interval        Collection interval (default 1s)
--log-level       Log level: debug, info, warn, error
--log-format      Log format: text, json
--consensus-addr  Consensus listen address
--peers           Comma-separated peer addresses

Architecture

…

API

Metrics

Prediction:

• sentinel_prediction_failure_probability - 0 to 1
• sentinel_prediction_confidence - model confidence
• sentinel_prediction_time_to_failure_seconds - estimated TTF
• sentinel_prediction_preemptive_migrations_total - migration count

Partition:

• sentinel_partition_detected - 0 or 1
• sentinel_partition_duration_seconds - how long partitioned
• sentinel_consensus_is_leader - leader status
• sentinel_partition_autonomous_decisions_total - decisions made offline

Rate limiting:

• sentinel_ratelimit_dropped_total - messages dropped by rate limiter
• sentinel_ratelimit_enabled - rate limiter state

Key Features

Circuit Breaker - Protects against cascading failures when the K8s API is unreachable. Configurable thresholds with automatic recovery.

Exponential Backoff - Peer reconnection uses backoff with jitter to prevent thundering herd during network recovery.

Rate Limiting - Token bucket rate limiter on consensus messages to handle bursty traffic.

Structured Logging - JSON or text output with log levels, component tags, and context propagation.

Health Checks - Kubernetes-native liveness and readiness probes with detailed component status.

Development

make test          # Run tests
make build         # Build for current platform
make build-arm64   # Build for ARM64 edge devices
make docker-build  # Build container image

Requires Go 1.21+.

License

MIT