Most PXE tutorials assume a flat lab network: one DHCP server, one TFTP root, MAC addresses you already know, and an image that installs itself. None of those assumptions survive a production rack.
Here are three failure modes we kept hitting, and the design changes that made them stop hurting.
1. Blind ARP sniffing to identify the machine that just booted
The classic approach listens for ARP to map IP to MAC, then matches that MAC against the IPMI/BMC entry so the tool can cut power and boot from network.
It breaks the moment your provisioning network and your IPMI network are physically isolated — which is how most operators run them. And when a heuristic binding guesses wrong, you power-cycle a production host.
What works better: let iPXE read the SMBIOS fingerprint and drop the host into a pending pool. Nothing gets power-controlled until a human confirms the physical label, serial or FRU. Read-only until confirmed.
2. Windows driver injection, and the password nobody can reset
Linux provisioning is mostly a solved problem: preseed, autoinstall, kickstart. Windows is where unattended builds fall over — a missing storage driver during WinPE setup, a UEFI/GPT partition script that is one character off, or a local admin password nobody wrote down.
Two changes help:
- Keep driver sets attached to the image template, not to the operator's laptop, so a reinstall is reproducible.
- Make offline password recovery a first-class operation. If the only recovery path is "wipe the disk", operators stop using your tooling for anything risky.
3. Answer files served in cleartext
The answer file usually contains credentials, and it is usually fetched over TFTP on the same LAN everybody can reach. Spoof a MAC and you can pull it.
Design around it:
- short-lived task tokens, scoped to one machine and one task
- an encrypted template vault rather than plain files on disk
- signing on any command that can cut power or wipe a disk
The checklist we ended up with
- Discovery is read-only; power actions need explicit confirmation.
- Verify SMBIOS / serial before any destructive step.
- Cache images locally so a dead uplink does not stop a build.
- Keep task state at the edge — a broken link should not lose a half-built machine.
- Prefer one binary over a five-container stack you have to babysit at 3am.
None of this is exotic. It is just the difference between a demo and something an operator will trust with a rack.
Disclosure: I work on PXE Commercial, a self-hosted bare-metal lifecycle automation engine built around these constraints — iPXE discovery, unattended Windows/Linux provisioning, offline recovery, multi-site orchestration, single Go binary. The lessons above apply whichever tool you use.
Top comments (0)