posts

Jul 20, 2020

Deploying Elixir/Phoenix Release to Production

Estimated Reading Time: 12 minutes (2535 words)

Updates (12th August 2020): Fix my mistake on using bin/app restart.

Updates (7th October 2020): Include the reason why bin/app restart doesn’t work

In my previous post “Building Elixir/Phoenix Release With Docker”, I wrote about how I build Elixir release with Docker and extract the tarball. Today, I am going to share how I deploy Elixir release to the production server.

Do note that, the way I deploy works best for hobby or small projects. For larger scale system, consider using other tools.

For the sake of simplicity, this post assume that your remote server has reverse proxy like nginx setup and pointing port 80 towards your application port 4000. If you depend on the database, it’s assumed that the database is up and running.

Steps for initial release

Before we start, let’s briefly talk about the steps involved to deploy our release:

  1. Copy the release tarball to the remote server.
  2. Extract the tarball on the remote server.
  3. Start your application by running /bin/app_name daemon

This is equivalent to the following bash script:

#!/bin/bash

set -e

# Variables
APP_NAME="$(grep 'app:' mix.exs | sed -e 's/\[//g' -e 's/ //g' -e 's/app://' -e 's/[:,]//g')"
APP_VSN="$(grep 'version:' mix.exs | cut -d '"' -f2)"
TAR_FILENAME=${APP_NAME}-${APP_VSN}.tar.gz

# Replace with your remote server IP address or hostname
HOST="192.162.1.1"

# Create directory for our app first. In this case we are creating
# the folder at our user home directory. WHICH MIGHT NOT BE THE BEST PRACTICE.
ssh $HOST mkdir -p $APP_NAME/releases

# Use scp to copy our tarfile from local machine to remote server
# We are copying the tarfile to the directory we created above.
scp $TAR_FILENAME $HOST:~/$APP_NAME/releases/$TAR_FILENAME

# Extract the tarfile
ssh $HOST tar -xzf $APP_NAME/releases/$TAR_FILENAME -C $APP_NAME

# Source environment variable and start our Elixir application
ssh $HOST "source ~/$APP_NAME/.env  && ~/$APP_NAME/bin/$APP_NAME daemon"

# Remove tarfile that is copied in the 2nd step.
ssh $HOST rm "~/$APP_NAME/releases/$TAR_FILENAME"

Here we used some of the common command such as:

If you save this file as ./deploy in your application root directory locally (or where your tarball is available) and run chmod +x ./deploy, you should be able to deploy your initial release by simply running ./deploy. Simple and straightforward right?

Side Topic: SSH Tips and Tricks

Notice the pattern we use here in ssh $HOST <command to run>. We are essentially running the command on our remote server by first ssh into the server and executing the command.

If you’re new to this, go ahead and run ssh <user>@<ip> "ls -la" on your local machine. You should be able to see the same result as running ls -la in your remote server.

If you frequently ssh to a particular IP, you can add the following to your ~/.ssh/config:

Host prod-server # Host Name
  user kai # SSH as user kai
  Hostname 192.168.1.1 # IP address

With this configuration, you can now directly ssh prod-server instead of using ssh [email protected].

Steps for updating subsequent release

Subsequent release involves the similar steps as the above. The difference is before starting the new version, we need to stop our old version server first. However, there is a couple of things that is good to know before we proceed.

Stopping application take some times

Script like this won’t work:

bin/app stop
bin/app start

This is because it takes time for the old application to shutdown gracefully. Hence, running start command immediately would likely to cause the following error:

Protocol 'inet_tcp': the name app@hostname seems to be in use by another Erlang node

To overcome this issue, we need to either:

Why can't we just use bin/app restart?

In the previous version of this post, we are using bin/app restart. However, after I use the code personally in one of my project, I realized that by using `bin/app restart`, the application will not be started as the latest version.

It is restarted as the previous version of the application, which is not what we expected when we want to update our application right. Hence, we will need to write some custom logic as mentioned above.

I haven't figure out why it behave like this. If you happen to know, please let me know! If I manage to find out why, I'll update this part once again.

One of the reader commented on the commit of this post to include the reason why it isn't working as expected. This is because restart normally send HUP signal to the running code and this does not stop the code like KILL signal does. Hence, that's why we are still getting the old version.

Replacing old release with new release will cause bin/app command not working expectedly

That’s not the only thing we need to overcome. Another tricky one would be, if we were extract our new release tarball, which then replace our old release, bin/app pid would not work as expected at the time of writing (11th July, 2020) without any additional configuration. You’ll get the following error instead:

--rpc-eval : RPC failed with reason :nodedown

Why would this occur? After some experimentation, I have found out that this happen because every time we build a new release in Docker, as we run mix release, it would generate a new random cookie.

Which is then written to our releases/COOKIE file, and if we don’t have RELEASE_COOKIE set in our environment, this cookie would be used.

Every time we extract our new release tarball, the releases/COOKIE might be updated to a different cookie and cause the command unable to talk to our running application.

Here’s what the Erlang documentation mentioned:

When a node tries to connect to another node, the magic cookies are compared. If they do not match, the connected node rejects the connection.

Our new release cookie doesn’t match with our old release (running application) cookie. Thus, they are not able to talk to each other.

But, how is it related to our bin/app pid command? Isn’t it just a normal command that get the process id of the application?

It is, but internally, the command is using rpc mechanism to talk to the node, which spin up a hidden node and evaluate some code on the remote node (our running application) (refer to elixir --help, Distribution options for more details).

Here is the command executed underneath every time we run bin/app pid:

/home/kai/app/releases/0.1.1/elixir --hidden --cookie COOKIE --sname rpc-29e0-app --boot /home/kai/app/releases/0.1.1/start_clean --boot-var RELEASE_LIB /home/kai/app/lib --rpc-eval app IO.puts System.pid()

Side Note: How can we know the command running underneath?

An easy way to know what’s the command running underneath of a executable script is adding set -x on top of the script file. Instead of having the original bin/app script that looks like this:

#!/bin/sh
set -e

SELF=$(readlink "$0" || true)
if [ -z "$SELF" ]; then SELF="$0"; fi
RELEASE_ROOT="$(cd "$(dirname "$SELF")/.." && pwd -P)"
...

We can modified the script to get more details by adding a single character:

#!/bin/sh
# Just add extra x here
set -ex

SELF=$(readlink "$0" || true)
if [ -z "$SELF" ]; then SELF="$0"; fi
RELEASE_ROOT="$(cd "$(dirname "$SELF")/.." && pwd -P)"
...

With a simple change, now every time when we execute the bin/app command, a detailed log will be output to show what is being run underneath.

I actually came across this while going through Buildkite documentation for writing build scripts. Go ahead and read about it if you’re interested in the details behind.

Solution

The first thing we need to resolve is to ensure that every time we start our release, the same cookie is used. Fortunately, this can be easily done by using RELEASE_COOKIE environment variable or putting the cookie in our release configuration in mix.exs:

def project do
  [
    app: :app_name,
    ...
    releases: [
      app_name: [
        cookie: "<YOUR COOKIE>",
        steps: [:assemble, :tar]
      ]
    ]
  ]
end

The documentation recommend to use a long and randomly generated string for your cookie, which can be generated using the following code:

Base.url_encode64(:crypto.strong_rand_bytes(40))

Alternatively, you can use the RELEASE_COOKIE environment variable. In my case, it would be placing it in my local .env.production file:

export RELEASE_COOKIE=<YOUR COOKIE>

where later on, I have the following command that copy it to the remote machine as my environment variable file .env:

# scp <source> <host>:<destination>
scp .env.production $HOST:~/$APP_NAME/.env

2. Add script to deploy new release

The only difference between the script for initial release and subsequent release is the part where we start the application:

So, instead of:

source ~/$APP_NAME/.env  && ~/$APP_NAME/bin/$APP_NAME daemon

This is how it looks like:

# ==========================================
# Copying .env.production to remote as .env
# ==========================================
scp .env.production $HOST:~/$APP_NAME/.env

# =============================
# Copying new release to remote
# =============================
scp $TAR_FILENAME $HOST:~/$APP_NAME/releases/$TAR_FILENAME
ssh $HOST tar -xzf $APP_NAME/releases/$TAR_FILENAME -C $APP_NAME/

# ===================
# Start to trap error
# ===================

# This is because bin/app stop will return error
# if the application is not running.

# Furthermore, we want to get the status code of
# bin/app pid.
set +e

# ========================
# Stop running application
# ========================
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME stop"

# ================================
# Check if application has stopped
# ================================
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME pid"

# if getting process id of application return error
# it means that the application has been stopped
while [ $? -ne 1 ]
do
  ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME pid"
done

# =================
# Start application
# =================
# Starting the application in daemon mode
ssh $HOST "source $APP_NAME/.env  && $APP_NAME/bin/$APP_NAME daemon"

# ========================
# Health Check Application
# ========================

# Repeatly use rpc to talk to our node until it succeed, which indicate our node
# is now up and running.
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME rpc 'IO.puts(\"health-check\")'"
while [ $? -ne 0 ]
do
  ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME rpc 'IO.puts(\"health-check\")'"
done

# ===================
# Stop trapping error
# ===================
set -e

These code basically done the following through ssh:

This is good enough if you have only one production server. If you have more than one, consider looping through the IP addresses and extract the code into function.

You might be wondering why instead of using bin/app eval IO.puts("health-check"), we use rpc. This is because eval do not communicate with the node to execute the code. Hence, even if the node is not up yet, the execution will still be successful.

Side Note: Not the best way to health check

This is not the best way to health check your application. For the following reasons:

Glue it all together

To sum up, this is the bash script ./deploy that I used for deploying initial or subsequent release of my side projects:

#!/bin/bash

set -e

APP_NAME="$(grep 'app:' mix.exs | sed -e 's/\[//g' -e 's/ //g' -e 's/app://' -e 's/[:,]//g')"
APP_VSN="$(grep 'version:' mix.exs | cut -d '"' -f2)"
TAR_FILENAME=${APP_NAME}-${APP_VSN}.tar.gz
HOST="do"

bold_echo() {
  echo -e "\033[1m---> $1\033[0m"
}

bold_echo "Creating directory if not exist..."
ssh $HOST mkdir -p $APP_NAME/releases/$APP_VSN

bold_echo "Copying environment variables..."
scp .env.production $HOST:~/$APP_NAME/.env

bold_echo "Copying release to remote..."
scp $TAR_FILENAME $HOST:~/$APP_NAME/releases/$TAR_FILENAME
ssh $HOST tar -xzf $APP_NAME/releases/$TAR_FILENAME -C $APP_NAME/

set +e
bold_echo "Waiting for existing application to stop..."
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME stop"
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME pid"

# if getting process id of application return error
# it means that the application has been stopped
while [ $? -ne 1 ]
do
  ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME pid"
done

bold_echo "Starting application in daemon mode..."
ssh $HOST "source $APP_NAME/.env  && $APP_NAME/bin/$APP_NAME daemon"

bold_echo "Health checking application..."
# Waiting for application to start
ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME rpc 'IO.puts(\"health-check\")'"
while [ $? -ne 0 ]
do
  ssh $HOST "source $APP_NAME/.env && $APP_NAME/bin/$APP_NAME rpc 'IO.puts(\"health-check\")'"
done
set -e

bold_echo "Application started!"

bold_echo "Removing remote tar file..."
ssh $HOST rm "~/$APP_NAME/releases/$TAR_FILENAME"

bold_echo "Removing local tar file..."
rm $TAR_FILENAME

Don’t forget to make it executable by running chmod +x ./deploy, and now you can deploy your Elixir release by running ./deploy locally.

Here, I didn’t cover how I run my migration but a quick way is just adding the following command after we start the application (assuming you have added the code mentioned in the Phoenix Release Documentation):

ssh $HOST "source ~/$APP_NAME/.env && ~/$APP_NAME/bin/$APP_NAME eval 'App.Release.migrate()'"

The reason we need to source the .env is because our runtime require some environment variable to be available in order to execute it.

Also, some additional notes on the extra stuff added in the scripts:

Wrap Up

That’s all. Building and deploying Elixir release can be simple once you know the building blocks. However, do remember that this might not be the best approach to deploy. It really depends on your context. For my personal projects, I found it to be sufficient as I only have a single production server and I am the only one who deployed it.

But is that all for my deployment process for Elixir/Phoenix release? Of course not! The next one I would share in the future would be deploying our release using Blue Green Deployment strategy with nginx. So do stay tuned!

Update (15 May 2021): I have finally completed my writing on deploying Elixir/Phoenix release using Blue Green Deployment with Nginx. If you are interested, continue to read it here.