Netdata — Real-Time Infrastructure Monitoring & Observability

What Netdata Does

• Auto-Discovery: Automatically detects and monitors OS, containers, databases, web servers, and 800+ integrations
• Per-Second Metrics: Collects metrics every second (not every 15s like Prometheus) for real-time visibility
• Zero Config: Install and immediately see 2,000+ metrics — no YAML files, no exporters to deploy
• ML-Powered Alerts: Machine learning detects anomalies in every metric automatically
• Beautiful Dashboards: Interactive, drill-down dashboards that update in real-time
• Distributed Architecture: Deploy agents everywhere, view all data in one place via Netdata Cloud
• Low Overhead: ~1% CPU, ~100MB RAM for monitoring an entire server with thousands of metrics
• Long-Term Storage: Built-in tiered storage with configurable retention

Architecture

┌─────────────────────────────────────────────┐
│  Netdata Agent (on each server)             │
│  ┌───────────┐ ┌──────────┐ ┌────────────┐ │
│  │Collectors │ │ ML Engine│ │ Dashboard  │ │
│  │(800+ auto)│ │(Anomaly) │ │ (Built-in) │ │
│  └───────────┘ └──────────┘ └────────────┘ │
│  ┌───────────┐ ┌──────────┐ ┌────────────┐ │
│  │ TSDB      │ │ Alerts   │ │ Streaming  │ │
│  │(Per-second)│ │(ML+Rules)│ │ (to Cloud) │ │
│  └───────────┘ └──────────┘ └────────────┘ │
└─────────────────────────────────────────────┘

What Gets Monitored Automatically

System:
├── CPU (per core, per process, by type)
├── Memory (RAM, swap, page faults, NUMA)
├── Disk I/O (per device, latency, utilization)
├── Network (per interface, packets, errors)
├── Processes (count, states, context switches)
└── Sensors (temperature, fans, voltage)

Containers:
├── Docker (per container CPU, memory, I/O, network)
├── Kubernetes (pods, deployments, nodes)
└── LXC/LXD

Databases:
├── MySQL / MariaDB (queries, connections, replication)
├── PostgreSQL (locks, transactions, WAL)
├── Redis (commands, memory, keys)
├── MongoDB (operations, connections, replication)
└── Elasticsearch (indexing, search, cluster health)

Web Servers:
├── Nginx (requests, connections, status)
├── Apache (workers, requests, bandwidth)
├── HAProxy (frontend/backend, sessions)
└── Traefik (entrypoints, routers)

Applications:
├── Node.js, Python, Go, Java (runtime metrics)
├── RabbitMQ, Kafka (queues, messages)
├── DNS servers (queries, cache)
└── 800+ more integrations

Key Features

ML-Powered Anomaly Detection

Every metric gets a machine learning model trained on its historical patterns:

Normal: CPU usage follows daily work pattern
Alert:  CPU anomaly detected — usage 3σ above predicted

Normal: Disk I/O steady at 50 MB/s
Alert:  Disk I/O anomaly — unusual spike to 500 MB/s at 3am

No manual threshold configuration needed — ML learns what's normal for YOUR infrastructure.

Composite Charts

Drill down from high-level overview to individual metrics:

Server Overview → CPU → Per Core → Per Process → System Calls

Alert Notifications

# Built-in notification channels:
- Email (SMTP)
- Slack
- Discord
- PagerDuty
- Opsgenie
- Telegram
- Microsoft Teams
- Custom webhook

Streaming & Centralization

┌──────────┐     ┌──────────┐     ┌──────────┐
│ Agent 1  │────▶│          │     │ Netdata  │
│ (Web)    │     │  Parent  │────▶│ Cloud    │
│          │     │  Agent   │     │ (SaaS)   │
└──────────┘     │          │     └──────────┘
┌──────────┐     │          │
│ Agent 2  │────▶│          │
│ (DB)     │     └──────────┘
└──────────┘

Stream metrics from child agents to a parent for centralized dashboarding and long-term storage.

Netdata vs Alternatives

常见问题

Q: Netdata 和 Prometheus + Grafana 怎么选？ A: Netdata 适合快速部署和实时监控，开箱即用。Prometheus + Grafana 适合需要长期指标存储、自定义查询（PromQL）和定制化仪表盘的场景。两者可以共存——Netdata 导出指标到 Prometheus 也是常见架构。

Q: Netdata Cloud 是必须的吗？ A: 不是。每个 Netdata agent 都有完整的本地仪表盘。Cloud 是可选的 SaaS 服务，用于跨多服务器的统一视图。自托管用户可以用 parent agent 替代。

Q: 对服务器性能影响大吗？ A: 非常小。典型场景下 CPU 占用 ~1%，内存 ~100-150MB。Netdata 使用高效的 C 语言编写，专门优化了低开销采集。

来源与致谢

• GitHub: netdata/netdata — 78.4K+ ⭐ | GPL-3.0
• 官网: netdata.cloud