Skip to content

Kernel Telemetry

Real-Time Linux /proc Daemon & TimescaleDB Streamer

Context
Distributed Infrastructure Observability
Type
Web Architecture
Year
2025
Timeline
4 Weeks · <15MB RAM / <50ms Latency
Artwork for Kernel Telemetry

In plain language

Traditional server monitoring tools are like heavy generators that guzzle power just to tell you if the lights are on. We replaced that overhead with a featherweight native sensor that uses 95% less server memory, catches microsecond system crashes that normal monitors miss, and displays live server health on screen in under 50 milliseconds.

What it means for a business Reduces server monitoring memory overhead from 350MB to <15MB per node while giving operations teams sub-50ms instant visibility into live infrastructure bottlenecks.

Measured results

<15MB
Go Collector Resident RAM Footprint
50k+/s
Delta-Compressed Telemetry Events
<50ms
Kernel-to-Browser Stream Latency
82%
TimescaleDB Hypertable Compression
Legacy monitoring agents were consuming 350MB of RAM per server node and lagging 4 seconds behind real CPU and disk saturation spikes.

Diagnosing tail-latency spikes across distributed Linux servers is impossible when the monitoring tools themselves distort system behavior. Heavy agents built on Python or JVM runtimes routinely devour hundreds of megabytes of resident memory. Furthermore, polling REST endpoints every 5 seconds creates blind spots, missing sub-second CPU steal spikes, disk I/O bottlenecks, and network socket exhaustion.

A real-time Linux infrastructure observability system pairing a zero-allocation Go collection daemon with delta-compressed WebSockets, TimescaleDB time-series hypertables, and a 60fps HTML5 Canvas dashboard.

I developed a lightweight, statically linked Go daemon that reads the Linux /proc and sysfs virtual filesystems directly using pre-allocated byte pools, eliminating garbage collection pauses entirely. The daemon transmits binary delta frames over persistent WebSockets to an ingestion gateway backed by TimescaleDB hypertables. On the browser front end, telemetry frames bypass the DOM and stream directly into typed Float32Array buffers, flushing to hardware-accelerated Canvas charts once per requestAnimationFrame tick.

  • Zero-allocation Go daemon parsing Linux /proc and sysfs counters with reusable byte buffer pools at 100Hz
  • WebSocket backpressure controller with ring-buffer frame coalescing to prevent browser UI freezing under heavy bursts
  • Automated TimescaleDB continuous aggregates compressing historical telemetry by 82% while keeping live queries instant

Host memory consumption plummeted from 350MB to under 15MB of resident RAM per server node. Glass-to-glass latency from raw kernel counter to browser display dropped below 50ms under loads exceeding 50,000 telemetry events per second. Automated TimescaleDB compression policies shrank 90-day time-series data by 82% without degrading analytical query speeds.

How it was built, step by step.

How we replaced a 350MB lagging agent with a <15MB Go /proc daemon streaming 50k+ events/sec.
  1. Cols 01–03 · Days 1–4

    Zero-Allocation /proc & Ring-Buffer Specification

    Mapped Linux /proc/stat, /proc/meminfo, and /proc/net/dev byte layouts and set hard ceiling budgets: <15MB resident RAM and <50ms glass-to-glass stream latency.

    BUDGET SIGNED OFF<15MB RAM Target

  2. Cols 03–06 · Week 1–2

    Static Go Daemon & WebSocket Multiplexer

    Built the zero-dependency Go collector using pre-allocated byte buffers to eliminate GC pauses, streaming delta-compressed frames into TimescaleDB hypertables.

    DAEMON BENCHMARKED50,000+ Events/sec

  3. Cols 06–09 · Week 2–3

    Float32Array Reactive Console & 60fps Canvas Charts

    Connected incoming WebSocket binary frames directly into typed Float32Arrays in Nuxt, flushing dirty chart regions once per requestAnimationFrame tick.

    60FPS LOCKED<50ms Glass-to-Glass

  4. Cols 10–12 · Week 4

    Load-Test Soak & Dockerized Production Rollout

    Ran 72-hour continuous stress tests across containerized Linux nodes, verifying 82% TimescaleDB compression and zero memory leaks under spike load.

    SHIPPED TO PROD82% DB Compression

Stack

  • Go
  • TypeScript
  • WebSockets
  • TimescaleDB
  • Docker
  • UI/UX

Credits

Backend & Go Daemon — Norbert Brett (br3tt) · Storage — PostgreSQL & TimescaleDB · Frontend Console — Nuxt, TypeScript & HTML5 Canvas