Running on several machines

번역 준비 중

이 쪽은 아직 한국어로 옮기지 않았습니다. 아래는 영어 원문입니다.

One elanous works on one machine. When you have more than one — a laptop you work on, a desktop that stays on, a small cloud VM that faces the internet — you can join them into one fleet: every machine reports what it has and what it is for, one of them keeps the shared record, and that job can move to another machine when the first one sleeps or reboots.

This page sets that up from scratch. Every step is a command you can check right after you run it.

The two kinds of role

A machine can hold two kinds of role, and they behave differently.

	Seat	Duty
How many machines at once	exactly one	any number
Example	control — the machine that keeps the fleet record and accepts changes	workstation, compute, edge, character
What you write on each machine	whether it is a candidate, and its rank	the duties it takes on
What "primary" means	the candidate that holds the seat right now	—

• Seats move. You never say "this machine is the primary". You say "this machine is a candidate for the control seat, rank 1", and the fleet works out who holds it. The holder is the primary for that seat; the other candidates stand by.
• Duties describe what a machine is good for. They do not move and they can overlap. Suggested names:

Duty	Typical machine
workstation	the machine a person works at (browser, editor, terminal UI)
compute	always on, many cores — runs builds and parallel agent work
edge	a small cloud VM with a public address — webhooks, outside checks
character	hosts the browser screens of bot characters

You can use your own duty names (lowercase letters, digits and -).

A machine can be a candidate for a seat, take on duties, both, or neither. A small internet-facing VM, for example, is usually edge only and not a seat candidate — it should not become the machine that keeps your record.

Before you start

• elanous installed on every machine — see Install. Check with elanous --version on each.
• The machines can reach each other over a private network. We use Tailscale; any private network works. Do not put the control seat on a public address.
• Pick a short, lowercase id for each machine (home, studio, edge-1, …). It is how the fleet names the machine, whatever the operating system calls it.

1. Describe each machine

Run this on every machine, with that machine's own values:

# the laptop you work on — first candidate for the control seat
elanous machine set --id home --duty workstation --seat control:1

# an always-on desktop — second candidate, also does the heavy work
elanous machine set --id studio --duty compute --duty character --seat control:2

# a cloud VM — outside tasks only, never holds the seat
elanous machine set --id edge-1 --duty edge

Check:

elanous machine show

machine set only changes what you pass; run it again to add a duty or change a rank. The profile is stored in the elanous state folder with owner-only permissions.

2. Start the control seat on the rank-1 machine

On the rank-1 candidate (home above):

elanous control serve

It listens on 127.0.0.1:31413 and creates the fleet's admin tokens. Keep it running as a service (launchd on macOS, a systemd user unit on Linux) so it comes back after a reboot.

Make it reachable from the other machines over the private network only. With Tailscale:

tailscale serve --bg --https=8413 http://127.0.0.1:31413

The other machines will use https://<this-machine>.<your-tailnet>.ts.net:8413. The control seat accepts nothing without a token, so it is safe behind a private-network proxy — but never publish it on the open internet.

3. Join the other machines

Each machine gets its own token, and the token only lets that machine report about itself.

On the rank-1 machine, issue a token and hand it straight to the other machine through a pipe, so it is never shown on screen or saved in your shell history:

elanous control token issue studio \
  | ssh studio 'elanous control join --url https://home.<your-tailnet>.ts.net:8413 --machine studio --token-stdin'

On a machine you cannot reach with ssh, save the token to a file readable only by you and use --token-file <path> instead, then delete the file.

Check on the joined machine: it prints Joined <url> as <id>. On the rank-1 machine, elanous control token list shows the id and when it was issued. To cut a machine off, run elanous control token revoke <id>.

4. Report what each machine has

On each joined machine, run the member and tell it what to report:

elanous control member --resource image:registry=http://127.0.0.1:5050/v2/

It sends a heartbeat every 30 seconds with the machine's load, memory, duties and seat ranks, and checks that each resource you listed answers. Run it as a service too.

See the whole fleet from any machine:

elanous resources list
elanous resources where registry     # one resource by name, or every resource of a kind

5. Let the control seat move (optional)

Up to here the rank-1 machine always holds the control seat. If you want the seat to move to the next candidate when that machine sleeps, reboots or fails, the candidates need one shared object they can all read and write — today a Google Cloud Storage bucket.

On every seat candidate:

elanous config set roles.bucket '"gs://<your-bucket>"'
elanous role watch

Run role watch as a service. Every candidate checks the seat once a minute:

• Nobody holds it — rank 1 takes it at once; rank 2 only after it has seen the seat empty for a while; rank 3 later still. No machine needs a list of the others.
• The holder stops renewing — the next candidate takes over after a grace period, lower ranks first.
• The holder comes back — it does not take the seat back by itself. That avoids the seat bouncing each time a laptop wakes up.

Then start the control seat on every candidate with the lease turned on, so only the current holder accepts changes:

elanous control serve --follow-lease

Check who holds the seat, from any candidate:

elanous role status

Planned reboots

Hand the seat over first, then reboot:

elanous role handoff --to studio     # on the current holder
elanous role accept                  # on studio, if it does not accept on its own
# … reboot the first machine …
elanous role handoff --to home       # when it is back, hand the seat back

Doing this for each machine in turn lets you update every machine without the seat ever being empty.

Troubleshooting

You see	Do
invalid machine	The id must be lowercase letters, digits or -. Set it with elanous machine set --id <id>. elanous role whoami shows the name in use and where it came from.
roles.bucket error from role watch	Step 5 needs a shared bucket — set roles.bucket, or skip step 5.
unauthorized when joining	The token was cut or mistyped. Issue a new one (control token issue <id> replaces the old one) and pipe it again.
A machine is missing from resources list	Its member is not running, or cannot reach the control seat's address. Run elanous control member --once --json on it to see one report.
role status shows nobody after starting	No rank-1 candidate is running role watch. Check elanous machine show on each candidate.