Bare-Metal Distributed Browser Grid & HAProxy Load Router

Senior Automation Engineer · Year 1 · Q2 (Previous SaaS)

Platform & InfrastructureYear 1 · Q2 (Previous SaaS)

Built an auto-scaling, bare-metal browser grid handling 10,000 parallel test threads with custom Linux cgroup memory fencing and HAProxy session routing.

⚡10,000 Parallel Threads Sustained
⚡85% Infrastructure Cost Savings vs. Cloud Vendors
⚡Sub-500ms Browser Instance Provisioning
PythonSelenium NodeHAProxyLinux cgroupsBash

Situation

Outsourced browser cloud providers introduced high latency, network timeouts, and annual licensing costs exceeding $120,000. Additionally, parallel execution nodes on shared VMs frequently crashed due to unmonitored Chrome memory leaks consuming system Swap space.

Task

Architect an in-house bare-metal browser grid infrastructure capable of managing 10,000 concurrent browser sessions with strict hardware process isolation and zero cross-session session state leakage.

Action

  • HAProxy Dynamic Routing Layer: Configured HAProxy as an ingress load balancer that analyzed Incoming WebDriver HTTP commands and routed session IDs (/session/{id}) dynamically to healthy downstream Linux worker nodes.
  • Kernel Resource Fencing: Implemented Linux cgroups to restrict each headless browser process to 512MB RAM and 1 CPU core, forcibly terminating rogue processes before they starved neighboring execution threads.
  • Ephemeral Xvfb Instance Recycling: Wrote a background daemon in Python that killed and re-spawned virtual framebuffer displays (Xvfb :display_id) every 50 sessions to reclaim leaked memory buffers.

Result

  • Saved $120,000/year in third-party vendor licensing costs.
  • Achieved 99.8% session initiation reliability across 10,000 parallel test runs.