Launch compute engine instance when connecting via SSH

launch-gc-vm.sh

#!/bin/bash
set -x

# Get command options
export project="$1"
export zone="$2"
export vm="$3"
echo "Launching instance $project/$zone/$vm ..."

# Alias with project/zone options
gcloud="gcloud --project=$project"

# Check current status first
export status="$($gcloud compute instances describe --zone=$zone --format="value(status)" $vm)"
echo "$project/$zone/$vm: $status"

# Launch if not running yet
if [ "$status" != "RUNNING" ]; then
    $gcloud compute instances start --zone=$zone $vm || true
fi

# Wait to boot
count=0
while ! $gcloud compute instances get-serial-port-output --zone=$zone $vm | grep 'Startup finished in' ; do
	sleep 1
	count=$(( count + 1 ))
	if [ $count -gt 30 ] ; then
		echo "Timed out after $count attempts"
		exit 1
	fi
done
echo "Instance launched."

ssh-config

# Match and execute the launch script, so when connecting it will launch VM if it's not running
Match host myvm exec "/path/to/launch-gc-vm.sh myproject us-central1-f myvm"

# Use the IP address or DNS name if you have a public entry for the target host
# Also, you can use the SSH key that `gcloud compute ssh myvm` will generate for
Host myvm
	HostName X.X.X.X
	User myuser
	IdentityFile /home/myuser/.ssh/google_compute_engine

Create a ~/.ssh/config entry using the Match directive, and point it to the script. When issuing

ssh myvm

The script will launch the VM and wait for it to be running with the system booted before executing SSH. Pro-tip: add this to your Visual Studio Code Remote SSH and get an on-demand setup with an automated shutdown at late night!

PS: the "Wait to boot" part of the script is tested with Debian systemd log to serial output; adjust it to match your needs (or comment out if your boot sequence is fast; usually 30s is enought to boot on GCP, so a basic sleep 30 will be fine).