scuttle handles SIGTERM or spot termination notices by optionally draining or deleting the current Kubernetes node. It's best run as a systemd unit designed to gracefully stop on shutdown (shown below).
- Uncordon the node on start (reboot case)
- Handle SIGTERM (unit stop or shutdown)
- Monitor instance metadata for termination notices (AWS, Azure planned)
- Drain and/or delete (de-register) a Kubernetes node
- Evict Pods left by Kubelet's GracefulNodeShutdown
scuttle compliments Kubelet's GracefulNodeShutdown feature, which only handles a part of gracefully stopping the Kubelet. See the blog post to learn more. In effect, scuttle is just a Go implementation of the bash scripts shown in the posts.
scuttle must be run in a systemd unit that is designed to stop before shutdown. Systemd shutdown can be subtle. Read systemd Shutdown Units and Kubelet Graceful Shutdown for background.
[Unit]
Description=Scuttle Kubelet before Shutdown
After=multi-user.target
[Service]
Type=simple
ExecStartPre=-/usr/bin/podman rm scuttle
ExecStart=/usr/bin/podman run \
--name scuttle \
--network host \
--log-driver=k8s-file \
--env KUBECONFIG=/var/lib/kubelet/kubeconfig \
-v /var/lib/kubelet:/var/lib/kubelet:ro,z \
--stop-timeout=60 \
quay.io/poseidon/scuttle:v0.1.0 \
-platform=aws
ExecStop=/usr/bin/podman stop scuttle
TimeoutStopSec=180
SuccessExitStatus=143
[Install]
WantedBy=multi-user.target
Users of Fedora CoreOS or Flatcar Linux should use the Butane Config:
variant: fcosversion: 1.4.0systemd:
units:
- name: scuttle.servicecontents: | [Unit] Description=Scuttle Kubelet before Shutdown After=multi-user.target [Service] Type=simple ExecStartPre=-/usr/bin/podman rm scuttle ExecStart=/usr/bin/podman run \ --name scuttle \ --network host \ --log-driver=k8s-file \ --env KUBECONFIG=/var/lib/kubelet/kubeconfig \ -v /var/lib/kubelet:/var/lib/kubelet:ro,z \ --stop-timeout=60 \ quay.io/poseidon/scuttle:v0.1.0 \ -platform=aws ExecStop=/usr/bin/podman stop scuttle TimeoutStopSec=180 SuccessExitStatus=143 [Install] WantedBy=multi-user.target - name: scuttle.pathenabled: truecontents: | [Unit] Description=Watch for Kubelet kubeconfig [Path] PathExists=/var/lib/kubelet/kubeconfig [Install] WantedBy=multi-user.targetConfigure via flags.
| flag | description | default |
|---|---|---|
| -node | Kubernetes node name | $HOSTNAME |
| -platform | Platform to poll for termination notices | none |
| -uncordon | Uncordon node on start | true |
| -drain | Drain node on stop | true |
| -delete | Delete node on stop | true |
| -channel-id | Slack Channel ID | "" |
| -token | Slack Bot Token | "" |
| -webhook | Slack Webhook URL | "" |
| -log-level | Logger level | info |
| -version | Show version | NA |
| -help | Show help | NA |
Other values are set via environment variables.
| variable | description | default |
|---|---|---|
| KUBECONFIG | Path to Kubeconfig | "" |
| HOSTNAME | Current node name | "" |
AWS and Azure provide warnings via instance metadata (2 min) before spot terminations. scuttle can monitor platform specific instance metadata endpoints to trigger drain or delete actions before shutdown.
scuttle integrates with Slack to post shutdown lifecycle events. Stay informed about spot terminations, shutdowns, drains, deletions.
- Create a Slack App
A token mode can post messages, thread replies, and add reactions.
- Grant bot token scopes
chat:writeandreactions:writeto a channel via "OAuth & Permissions" - Get the bot token
- Invite the app to a channel (
/invite @mybot) and note the Channel ID
scuttle -channel-id C0FAKEFAKE -token bot-token
A webhook mode is available if you only have a Slack Webhook URL. However, it cannot thread replies or add reactions.
scuttle -webhook https://hooks.slack.com/...
To develop locally, build and run the executable.
Build the static binary or container image.
make build
make image
Run the executable.
export KUBECONFIG=some-dev-kubeconfig
export HOSTNAME=node-name
./bin/scuttle
Use Ctrl-C to emulate a node shutdown.
INFO[0000] main: watch for interrupt signals
INFO[0000] main: starting scuttle
INFO[0000] start scuttle hostname=ip-10-0-35-141
INFO[0000] scuttle: uncordon node hostname=ip-10-0-35-141
INFO[0000] drainer: uncordoning node node=ip-10-0-35-141
^C
INFO[0004] main: detected interrupt
INFO[0004] scuttle: stopping... hostname=ip-10-0-35-141
INFO[0004] scuttle: cordoning node hostname=ip-10-0-35-141
INFO[0004] drainer: cordoning node node=ip-10-0-35-141
INFO[0004] scuttle: draining node hostname=ip-10-0-35-141
INFO[0026] drainer: evicting pod node=ip-10-0-35-141 pod=redacted
INFO[0027] drainer: evicting pod node=ip-10-0-35-141 pod=redacted
...
INFO[0027] drainer: drained node node=ip-10-0-35-141
INFO[0027] scuttle: deleting node hostname=ip-10-0-35-141
INFO[0009] done