24 nodes, one control plane: loading the BitScope rack

The whole rack is on the control plane. Saturday afternoon the bench cluster hit 24 nodes online — a Raspberry Pi 5 control plane plus 23 Pi 4 compute nodes in a BitScope ER24A Edge Rack 24 — all running the 2026.07.2-dev.101 OS image. Enrolling the fleet took 44 minutes; powering it on took two.

The hardware:

BitScope ER24A Edge Rack 24: six CB04B Cluster Blades in rows A–F, four Pi slots per blade, one 19–24 V DC feed for the lot.
Every Pi in the rack has its own BMC wired to its serial port, bussed over RS-485 so one manager reaches all 24 — that part is a later devlog; today was getting the fleet enrolled and online.
Control plane: a single Pi 5 (node-9bbaa24a), provisioned 2026-07-07.
Rasputin currently caps a cluster at 24 nodes — a deliberate limit the UI is designed around. With the control plane off-rack on its own Pi 5, 23 of the rack’s 24 Pis enroll and one sits idle.

Timeline (2026-07-12, UTC, from the control-plane UI):

16:04 — control plane alone: NODES ONLINE 1/1, a hex grid of empty bays.
16:07–16:48 — 23 compute nodes enrolled, one every ~2 minutes. Each appears in the grid as PENDING — waiting… the moment its enrollment lands.
17:07:20 — rack power-on. Still 1/1; 23 hexes waiting.
17:07:48 — 21/21 online, 20 tasks in flight. Twenty nodes joined inside 28 seconds.
17:08:15 — 23/23.
17:09:21 — 24/24, task queue drained to 1. Done.

How enrollment works:

Rasputin provisions a cluster as a matched set: each node gets a join token bound to its node-id, and the control plane holds only hashes of the set. A stolen SD card doesn’t impersonate a neighbor; an unenrolled board doesn’t join at all.
Each node’s 64 GB SD card is flashed individually from the control-plane interface — the image plus that node’s seed in one write — then slotted into its bay. No console cable, no per-node config. The ~2-minute cadence in the timeline is the human loop, not the software.
On power-on the node finds the control plane, presents its token, and goes PENDING → ONLINE on its own. Zero keystrokes between wall switch and 24/24.

The numbers worth keeping:

Power-on to all 24 online: 121 seconds, most of it absorbed in the first half-minute.
Control plane load after the join storm: 1% CPU, 3% memory — on a Pi 5, running agent v2026.07.2-dev.47.
Tasks spiked to 20 during the storm and drained without intervention.

Takeaway: enrollment is human-paced — 44 minutes of flashing and slotting SD cards. The software side of a 24-node bring-up took 121 seconds and left the control plane at 1% CPU.

There’s a 21-second time-lapse of the grid filling up from 1 node to 24 at the top of this post — the UTC clock in the header ticking through the real hour is the only editing trick.

Rasputin is an open-source (AGPL) homelab cluster system — a control plane, node OS, and firewall image that make a few Raspberry Pis or N100 boxes behave like one appliance. It’s pre-alpha, on GitHub, and I’m looking for a handful of design partners to run it and tell me what’s broken.