// the lab

Live experiments.

Each tile is one of my projects, deployed and embedded right here. Open one and click around. It's the real app, not a recording.

// experiments 12 live
EXP_01 · gpu-fleet LIVE

gpu-fleet

Control plane for bare-metal GPU clouds: node lifecycle over Redfish, topology-aware capacity reservations, tenant attestation and automated remediation. Go control plane, Rust node agent, Postgres. A chaos service keeps a 50-node simulated fleet breaking and healing around the clock on a live, read-only demo.

  • Node lifecycle as a pure state machine, with Postgres CHECK constraints repeating its invariants
  • Provisioner workers claim nodes with SKIP LOCKED leases; fenced results stop a stalled worker overwriting its successor
  • Topology-aware reservations (same_rack, same_pod, any) with best-fit placement that keeps whole racks free
  • Exclusion constraint on (node_id, period) makes double booking impossible in the database itself
STACK
Go · Rust · PostgreSQL · Backend · Infrastructure
HOST
isidorsson.github.io
TYPE
Backend