Running a Spark cluster fully air-gapped — and the traps that come with it
Field note · from operating air-gapped GB10 clusters
Plenty of GB10 owners need the cluster off the internet entirely — regulated data, sensitive models, or just principle. The forums have an open, unanswered thread asking exactly how to update an air-gapped Spark. It's doable, but the platform assumes connectivity in several places, and each assumption is a trap when you cut the cord.
Trap 1: the update path assumes internet
The mandatory first-boot update and subsequent driver/firmware updates are built around pulling from the network — and getting them onto an offline box, in the right order, without bricking the machine, is genuinely fiddly — bricked-on-update reports are common even for connected machines. There's an offline procedure, but the sequencing and the staging matter, and a wrong step here costs you a node.
Trap 2: /etc/hosts is not an air gap
The most dangerous false sense of security. Blocking domains in /etc/hosts feels like isolation, but plenty of components resolve around it, and some telemetry paths don't use the hostnames you blocked. A real air gap is verified with a network capture showing zero egress, not a config file you hope is complete. We've caught inference stacks phoning out fingerprints from machines their operators believed were sealed.
The rule: "air-gapped" is a property you measure at the wire, not a checkbox you set in software. If you can't show a capture with no outbound packets under load, you don't yet know what your stack is sending.
Trap 3: the serving stack wants to reach out too
- Model and tokenizer loading often defaults to fetching from a hub; offline mode has to be forced correctly or a launch silently tries to phone home and stalls.
- Telemetry and usage reporting are on by default in several inference stacks and take real work to fully silence — more than one flag, more than one component.
- Version pinning matters more offline: you can't just pull a fix, so the build you seal has to be the right one, verified, before it goes behind the gap.
We run air-gapped GB10 clusters ourselves, daily — it's not a special mode we read about, it's how we operate. That means the update sequence, the egress verification, and the offline serving config are things we've already solved the hard way rather than the theoretical way.
Need a Spark cluster that's genuinely sealed?
We operate air-gapped GB10 clusters as our default. We'll verify your egress at the wire, get the offline update and serving path working, and prove the seal with a capture — not a config file. Or it's free.
Book a 1-day diagnostic →Free to read and share. Numbers are from hardware we've measured directly; your versions and topology will differ — measure your own.