clusterctl

One binary. Every node. No surprises. clusterctl selects nodes with ClusterShell syntax, fans out commands, drives service processors, reinstalls nodes and administers Slurm, and asks before it changes anything.

$ clusterctl node select '@slurm:main&@rack:R02'
exe[0004-0009]

$ clusterctl exec -n '@slurm:main' --dedup -- uname -r
exe[0001-0009] (9): ok
  5.14.0-570.el9.x86_64
exe0010 (1): ok
  4.18.0-553.el8.x86_64

$ clusterctl slurm node drain 'ticket 4711: failing DIMM' -n exe0007
About to drain 1 host: exe0007
  reason: ticket 4711: failing DIMM
Continue? [y/N] y
drained exe0007

$ clusterctl provision reinstall -n '@rack:R02' --dry-run
Would reinstall 10 hosts: exe[0001-0010]
  everything on these machines is lost

With mise, which pins a version per project and verifies the download:

$ mise use -g github:GSI-HPC/clusterctl
$ clusterctl version

Or take the binary for your platform from the releases:

$ curl -fsSL -o clusterctl.tar.gz \
    https://github.com/GSI-HPC/clusterctl/releases/latest/download/clusterctl_linux_amd64.tar.gz
$ tar xzf clusterctl.tar.gz
$ install -m 0755 clusterctl ~/.local/bin/

Or build it:

$ go install github.com/GSI-HPC/clusterctl/cmd/clusterctl@latest

Go 1.26 or newer. The toolchain downloads itself if yours is older.

$ clusterctl config init --site lab --cluster alpha --domain hpc.example.org
$ $EDITOR ~/.config/clusterctl/*.yaml    # the files it listed; fill in what they ask
$ clusterctl config validate
$ clusterctl doctor
$ clusterctl node list

config init writes the least configuration that resolves, and only into an empty directory, and prints the path of every file it writes. On macOS that directory is ~/Library/Application Support/clusterctl, not ~/.config/clusterctl. examples/site/ is a complete configuration to take further settings from. The manual walks through both, and doc/migration.md maps every command and setting of the shell toolkit clusterctl replaces to its counterpart.

$ clusterctl completion bash > /etc/bash_completion.d/clusterctl
$ clusterctl completion zsh  > "${fpath[1]}/_clusterctl"

Node set completion offers the configured groups, and --context offers the configured contexts.

clusterctl's node set engine is a library of its own, go-nodeset, because the Go ecosystem had none:

import "github.com/GSI-HPC/go-nodeset"

ns, err := nodeset.Parse("exe[0001-0010]!exe0003")
fmt.Println(ns)          // exe[0001-0002,0004-0010]
fmt.Println(ns.Len())    // 9

Its language reference describes the language and the rules chosen where it differs from ClusterShell; doc/nodeset.md says what clusterctl adds to it.

The repository carries a mise.toml, so the toolchain comes from mise if you use it:

$ mise install  # Go, golangci-lint, sops and sind, at the versions CI uses

$ make test     # unit and command tests
$ make lint     # vet and formatting
$ make cover    # coverage
$ make build    # bin/clusterctl

The end-to-end tests run the binary against a Slurm cluster in Docker, which sind creates; they need Linux with Docker, and mise install installs sind:

$ make e2e-up   # create the cluster
$ make e2e      # run the tests against it
$ make e2e-down # delete it

The documentation site needs Hugo extended; see site/README.md.

Commits are Conventional Commits. Design notes and the decision records are in doc/; a change that takes a decision gets a record.

Copyright (C) 2026 GSI Helmholtz Centre for Heavy Ion Research GmbH http://www.gsi.de

LGPL-3.0-or-later. See COPYING and COPYING.LESSER.