I used to deploy my web apps on Ubuntu running on Digital Ocean but recently I switched to using CoreOS instead of Ubuntu.

For a while I didn’t understand CoreOS; a linux distro without package manager? How do I install more software on this thing?

Now I am a convert. CoreOS is not a Linux distro for end users. It’s a distro for deploying applications packaged as docker containers.

The benefit of using CoreOS is less configuration needed compared to e.g. Ubuntu.

I used to deploy multiple apps per server but for operational simplicity I moved to using one server per app. At $5 per server (my apps are written in Go, so they run comfortably on the smallest servers) it’s a reasonable cost.

Here’s my playbook for deploying an app on CoreOS. This example is how I deploy my blog.

1. Create a unique ssh key for the machine

For security it’s good to have a unique ssh key for each machine:

• ssh-keygen -t rsa -b 4096 -C "a comment" : creates a new ssh key, save it as id_rsa
• content of id_rsa.pub is what you give DigitalOcean as ssh key when creating a server

After server is created, verify you can login: ssh -i ./id_rsa core@<ip_address>.

2. Initial server setup

To make the scripts more re-usable, create ipaddr.sh:

# e.g. IPADDR=137.63.26.193
IPADDR=<ip address of the server>
# git likes to loose non-standard permissions. This is always called from
# scripts so a good place to ensure right permissions on id_rsa
chmod 0600 id_rsa

For convenience I also write login.sh:

#!/bin/bash
. ./ipaddr.sh
ssh -i ./id_rsa core@${IPADDR}

Usually a kernel benefits from one or more tweaks. Some tweaks don’t persist and have to be applied at startup. We’ll use systemd to run startup script and adjust kernel parameters. We need several files. files on the server:

startup.conf (/etc/sysctl.d/startup.conf on the server):

# http://security.stackexchange.com/questions/43205/nf-conntrack-table-full-dropping-packet
# https://coreos.com/os/docs/latest/other-settings.html
net.netfilter.nf_conntrack_max=131072
# in seconds, default value was 600 == 5 mins
net.netfilter.nf_conntrack_generic_timeout = 60
# in seconds, default value was 432000 i.e. 5 days
net.netfilter.nf_conntrack_tcp_timeout_established = 54000
# was 46992
fs.file-max = 131072

startup.conf contains the settings. Default values for connection tracking are so low that it’s easy for a malicious person to DoS your server just by opening a small number of connections to your http server.

This is for the smallest server with 512 MB of RAM. Increase for larger servers.

startup_script.sh (/etc/systemd/system/startup_script.sh on the server):

#!/bin/bash

# reload values from /etc/sysctl.d/startup.conf
sysctl --system
# make hashsize match net.netfilter.nf_conntrack_max (should be nf_conntrack_max / 4)
# https://security.stackexchange.com/questions/43205/nf-conntrack-table-full-dropping-packet
echo 32768 > /sys/module/nf_conntrack/parameters/hashsize

This script makes the changes. Now we need to make sure it gets called at startup.

startup.service (/etc/systemd/system/startup.service on the server):

# http://unix.stackexchange.com/questions/47695/how-to-write-startup-script-for-systemd
[Unit]
Description=things to do after each reboot
# just to be safe, run it after docker started
After=docker.service
Requires=docker.service

[Service]
Type=oneshot
# per https://www.digitalocean.com/community/tutorials/how-to-create-and-run-a-service-on-a-coreos-cluster
EnvironmentFile=/etc/environment
# '=-' means it can fail
ExecStart=-/etc/systemd/system/startup_script.sh

[Install]
WantedBy=multi-user.target

I have initial-server-setup.sh script to automate putting those files on the server and configure systemd to notice it.

#!/bin/bash
set -u -e -o pipefail

. ./ipaddr.sh

scp -i ./id_rsa ./startup.conf ./startup.service ./startup_script.sh core@${IPADDR}:/home/core/

ssh -i ./id_rsa core@${IPADDR} <<'ENDSSH'
sudo cp startup.conf /etc/sysctl.d/startup.conf
sudo chmod 0644 /etc/sysctl.d/startup.conf
sudo cp startup.service /etc/systemd/system
sudo cp startup_script.sh /etc/systemd/system
sudo chmod 0750 /etc/systemd/system/startup_script.sh
sudo systemctl daemon-reload
sudo systemctl enable /etc/systemd/system/startup.service

sudo bash -c "echo 32768 > /sys/module/nf_conntrack/parameters/hashsize"
sudo sysctl --system
rm startup.conf startup.service startup_script.sh
ENDSSH

Run initial-server-setup.sh and test that the changes are applied:

• login to the server with ./login.sh
• reboot with sudo shutdown -r now
• login again
• sudo sysctl -a | grep nf_conntrack_max

3. Use systemctld to automatically restart the app

When the server reboots we want the app to start automatically.

We also want the app to automatically restart if it crashes.

We’ll use systemd again as it comes with CoreOS.

Create blog.service (/etc/systemd/system/blog.service on the server) file which instructs systemd how to run a docker container named blog:

[Unit]
Description=blog
# this unit will only start after docker.service
After=docker.service
Requires=docker.service

[Service]
TimeoutStartSec=0
# per https://www.digitalocean.com/community/tutorials/how-to-create-and-run-a-service-on-a-coreos-cluster
EnvironmentFile=/etc/environment
# before starting make sure it doesn't exist
# '=-' means it can fail
ExecStartPre=-/usr/bin/docker rm blog
ExecStart=/usr/bin/docker run --rm -p 80:80 -v /data-blog:/data --name blog blog:latest
ExecStop=/usr/bin/docker stop blog
# restart if it crashes or is killed e.g. by oom
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

It’s a one-time operation but I still like to have it as a script named install-service.sh:

#!/bin/bash
set -u -e -o pipefail

. ./ipaddr.sh

scp -i ./id_rsa ./blog.service core@${IPADDR}:/home/core/blog.service

ssh -i ./id_rsa core@${IPADDR} <<'ENDSSH'
cd /home/core
sudo cp blog.service /etc/systemd/system
sudo systemctl enable /etc/systemd/system/blog.service
rm blog.service
ENDSSH

You need to re-run it after updating blog.service file.

4. Package the app as a Docker image and upload to the server

A script to build the app, package as Docker image and upload latest version to the server.

The most common advice for uploading/downloading docker images is to use docker registry. For simplicity I just use docker save/docker load and scp.

Here’s a build_and_upload.sh script:

#!/bin/bash

# build latest version of the app, upload to the server packaged as
# blog:latest docker image

set -u -e -o pipefail

. ./ipaddr.sh

dir=`pwd`
blog_dir=${GOPATH}/src/github.com/kjk/blog

echo "building"
cp config.json "${blog_dir}"
cd "${blog_dir}"
GOOS=linux GOARCH=amd64 go build -o blog_linux
docker build --no-cache --tag blog:latest .
rm blog_linux
cd "${dir}"

echo "docker save"
docker save blog:latest | bzip2 > blog-latest.tar.bz2
ls -lah blog-latest.tar.bz2

echo "uploading to the server"
scp -i ./id_rsa blog-latest.tar.bz2 core@${IPADDR}:/home/core/blog-latest.tar.bz2

echo "extracting on the server"
ssh -i ./id_rsa core@${IPADDR} <<'ENDSSH'
cd /home/core
bunzip2 --stdout blog-latest.tar.bz2 | docker load
rm blog-latest.tar.bz2
sudo systemctl restart blog
ENDSSH

rm -rf blog-latest.tar.bz2

You can see a sample Dockerfile and a sample docker_build.sh

After deploying the app for the first time you should restart the server (shutdown -r) to verify that the app will start up after reboot.

I also have a sceript logs.sh that logs in to the server and shows logs for the docker service in follow mode (like tail -f):

#!/bin/bash

. ./ipaddr.sh

ssh -i ./id_rsa core@${IPADDR} <<'ENDSSH'
cd /home/core
docker logs -f blog
ENDSSH