Node Reverse Proxy Deployment Via Gitolite, Upstart and Monit
This repo represents an attempt to detail an approach for automated deployment and hosting of Node applications on a remote server. I'll try to add step-by-step instructions to this readme file and where relevant commit some example scripts and config files too.
In order to kill two birds with one stone the example Node app we'll be deploying will be a reverse proxy listening on port 80, which will be useful in future when we want to run further services on the server on ports other than 80 but still reach them on clean URIs (e.g.
www.myapp.com as opposed to
- Node application code is source controlled under Git.
- The remote server is running Gitolite to enable collaborative development and deployment with multi-user access control to a remote Git repo.
- When new code is pushed to Gitolite a
post-receivehook is used to execute a shell script which moves the Node application files to their proper location on the server and restarts the app.
- Upstart and Monit are used to manage the Node application on the server, both for restarting on deployment and reboot and for displaying and reporting on status.
- Gitolite and Node apps will run under the
nodesystem users. The deployment script will be invoked as
- Server: Linode VPS running Ubuntu 10.04 Lucid
- Workstation: OSX Mountain Lion 10.8.1
- Node 0.8.9 (server)
- Node 0.8.6 (workstation)
- Git 18.104.22.168 (server)
- Git 22.214.171.124 (Apple Git-31.1) (workstation)
- Monit 5.0.3 (server)
- Gitolite 3.04-15-gaec8c71 (server)
- Upstart 0.6.5-8 (server)
Remote directory structures and file locations
I wasn't too sure about where to put the deployment script, live Node apps and related files. They could go in obscure traditional UNIX locations like
/var/opt/log but in the end I decided to group all Node stuff into
/var/node, aping how Apache sites often all go into
/var/www) and plumped for the following locations:
- Generic Node app deployment script:
- Live Node app:
- Live Node app's log file:
- Live Node app's pidfile:
Other salient locations include:
- Gitolite repo:
- The Git post-receive hook:
- The Upstart job config:
- Monit's config:
1. Setup Gitolite
Setting up Gitolite on the server is optional, but it makes it much easier to grant granular read/write access to your remote repo to coworkers. If you're pretty sure you're the only person who'll ever be working with the app then you could probably get away with setting up a bare Git repo yourself and working with it directly over SSH. You could possibly use GitHub too, but that's out of the question if you're hosting sensitive/private code and you don't want to pay for private repos.
Setting up Gitolite is beyond the scope of this document, but there's fairly good documentation here. Suffice to say I encountered a fair few hiccups while getting Gitolite to work, mainly revolving around SSH configuration, so I'm going to briefly talk about some of those sticking points and their remedies.
During installation Gitolite mandates the creation of a
gitolite) user on your server with limited privileges, and when you push code to Gitolite you do so over SSH and authenticate via public key. Gitolite works out whether you have access rights to the repo you're trying to push to by checking the public key you offer to the server, so needless to say it's important that SSH is presenting the correct one. I found the following command useful to verbosely debug a SSH connection that is mysteriously failing:
ssh -vT git@gitolite-host
It can also be useful to nail down which SSH credentials your system may be trying to use for a given user/key/host combination, in which case you could add an entry to your
~/.ssh/config file something like this:
Host gitolite-host HostName 0.0.0.0 IdentityFile ~/.ssh/id_rsa User git IdentitiesOnly yes
This example would define a
gitolite-host server alias pointing to your deployment server at IP
0.0.0.0 which would always connect as the remote user
git using the
~/.ssh/id_rsa key. The
IdentitiesOnly yes enforces the use of the specified key. OSX will sometimes cache a public key, especially when you've opted to save a key's password to the system keychain and/or
ssh-agent. So if you're really having trouble SSHing into Gitolite with right user/key/host combo you can purge that cache like so:
sudo ssh-add -L # Lists all public keys currently being cached sudo ssh-add -D # Deletes all cached public keys
2. The Node Application
I've added the Node reverse proxy application code to this repo here so you can look around the files. It uses nodejitsu's own node-http-proxy under the hood, which apparently is a robust proxy that sees a good deal of testing and production usage.
The app reads a JSON file specifying a set of proxy rules and starts listening on port 80. For a example, an incoming request to
www.my-node-app.com could be internally rerouted to a Node app running at
127.0.0.1:8022, a request to
www.my-domain.com could be proxied to an Apache vhost at
127.0.0.1:8000, and a request to
www.my-domain.com/api could be routed to a Node app sitting at
127.0.0.1:8023, and so on. Since this is Node, web sockets and arbitrary TCP/IP traffic will be proxied (hopefully) flawlessly. I think node-http-proxy also supports proxying of HTTPS over SSL/TLS too although I don't have that set up in the example app. As I understand it at the time of writing (Sept 2012) nginx and Apache via
mod_proxy still do not happily support web socket proxying out of the box.
The app is configured for server vs. dev environments via the environment variables
PORT. Later on you'll see the environment variables being exported to the Node app's process in the Upstart job configuration.
The app exposes a special route
/ping via custom middleware which we'll later see Monit use to periodically check the health of the proxy.
On the server we don't really want the app to run as root via
sudo however we're not allowed to bind to port 80 unless this is the case. One remedy would be to reroute all port 80 traffic to a higher port by reconfiguring the way IP traffic is internally routed. Apparently a command such as this routes packets coming in on port 80 to port 8000:
sudo iptables -t nat -A PREROUTING -p tcp --dport 80 -j REDIRECT --to-ports 8000`
But that sounds like easily forgotten extra configuration steps though, so in this case we'll be invoking our app as root initially but then immediately having it downgrade itself to the non-privileged user once internal setup has completed. Apparently this is how Apache handles listening on port 80.
At this stage it might be handy to create the non-privileged
node system user:
sudo adduser --system --shell /bin/bash --gecos "Node apps" --group --disabled-password --home /home/node node
3. The post-recieve hook
At this stage I'll assume you have some application code added to a fully working remote Gitolite repo on the server and are able to
pull to it with ease from your workstation.
The task of the Git
post-receive hook is to invoke a generic deploy script which moves the Node app files out of the bare Gitolite repo and into whichever location we decide to store our active node apps on the server whenever new code is pushed. First you need to SSH into the server and become Gitolite's
ssh user@host sudo su git
Navigate to the hooks folder for the relevant repo and start to edit the
post-receive file. It's important that this file is owned by and executable by the
git user, since that's the user that'll be executing it.
cd /home/git/repositories/proxy-node-app.git/hooks touch post-receive # If required, may already exist chmod u+x post-receive # If required, may already be executable nano post-receive
Make the contents look like (or similar to) the example
post-receive hook in this repo here.
In the hook we're attempting to invoke the
/var/node/node-deploy generic deployment script via a
sh command while passing down a configuration environment variable called
APP_NAME (we'll go on to make that script in the next section). Since this
post-receive hook will not be executing in an interactive shell it will bork at the
git user's attempt to
sudo, so the next thing we need to do is give the
git user the right to invoke
/var/node/node-deploy with the
sh command without a password.
Start to edit the
And add the following line at the bottom:
git ALL = (root) NOPASSWD: /bin/sh /var/node/node-deploy
(I think) this isn't as big a security concern as one might think since we're only giving the
git user password-less
sudo rights to this one particular script, which itself will only be editable and viewable by
We're not quite done with
/etc/sudoers though, we need to stop
sudo stripping out our
APP_NAME environment variable. Add the following at the top just above the
Defaults env_reset line:
Defaults env_keep += "APP_NAME"
Again, this shouldn't be too much of a security concern because we'll be handling the contents of
APP_NAME carefully in
You can see my version of sudoers here. It just has the default contents and the changes mentioned above.
Save and exit
/etc/sudoers. We now should be in a postion where we can push to our Gitolite repo and have the
post-receive execute, and having granted the
git user the right to invoke the deployment script as
root without asking for a password we shoud have the power to do any kind of filesystem manipulation we like. Now we need to write that script.
4. The generic deployment script
I'm calling this a "generic" deployment script because I'm aiming for it to be useful for publishing any reasonably non-complex Node app. To this end we use the
APP_NAME value passed in from the
post-receive hook to tailor the behaviour of the script. It doesn't have to be executable since we're invoking it via the
sh command. Go and ahead and create it:
cd /var/node sudo touch node-deploy
An example of this script is in this repo here.
The script above is fairly well commented so I won't go into much detail but basically it's simply syncronising the contents of the node app directory (in this case
/var/node/proxy-node-app/app) with the latest revision of the files in the bare Gitolite repo via
git checkout -f. Once that's been done it's changing the ownership of the files to the
node user and restarting the app via Monit.
You don't have to keep your node apps in
/var/node, anywhere will likely do, but after some research it seemed like a reasonably sensible location.
Upstart is an event driven daemon which handles the automatic starting of services at boot, as well as optionally respawning them if their associated process dies unexpectedly. Additionally it exposes a handy command line API for manipulating the services with commands like
sudo start servicename,
sudo stop servicename,
sudo restart servicename and
sudo status servicename etc.
A service can be added to Upstart by placing a valid Upstart job configuration file in the
/etc/init directory, with configuration files having the naming format
servicename.conf. It's important to note that the Upstart config files execute as
sudo shouldn't be needed in an Upstart job config.
In the context of a Node app we can use Upstart to daemonize the app into a system service. In other words we can start the app with a command like
sudo start proxy-node-app and have it run in the background without taking over our shell or quiting when we exit our SSH session. Upstart's
stop on and
respawn directives allow us to have the app automatically restarted on reboot and when it quits unexpectedly. What's more, Upstart's start/stop API provides a useful control point for Monit to hook onto (more on that later).
Upstart is already on your box if you're running Ubuntu 10.04 like me, but if you don't have it I think it's installable via
An example Upstart job configuration is in this repo here.
Once you have your job config file in place run the following command to have Upstart list out all its valid jobs. If yours is not in the list it means you have some sort of basic syntax error:
sudo initctl list
If all is well Upstart will now start and stop your Node app on shutdown and reboot, respawn it if it crashes, and allow you to manually control it via commands like
sudo start proxy-node-app. If your
start command fails to spin up the app as expected it means that some or all of the Upstart scripts are failing. One gotcha is that environment variables defined in an Upstart
env stanza do not expand when concatenated.
Monit is a utility for managing and monitoring all sorts of UNIX system resources (processes, web services, files, directories etc). We'll be using it to monitor the health of our proxy Node app, and indeed any other apps we decide to host on this box.
As mentioned above Upstart will automatically respawn services that bomb unexpectedly. However the system process that Upstart monitors could be alive and well but the underlying Node web app could be frozen and not responding to requests. Therefore a more reliable way of checking on our app's health is to actually make a HTTP request and this is where Monit comes in handy - we can set up Monit in such a way that unless it gets a
HTTP 200 OK response code back from a request to our app it will alert us and then attempt to restart it. This is why we added the special
/ping route to our app - it's a light weight response that Monit can hit that just returns the text
OK and a
So just to re-iterate: Upstart restarts the app on system reboot/crash and provides the start/stop command line API while Monit keeps tabs on its status while it's running and restarts it if it looks unhealthy. Monit is pretty powerful, and it can monitor all sorts of process metrics (uptime, cpu usage, memory usage etc.) but for our purposes we're just keeping it simple for now.
I installed Monit via
apt-get on Ubuntu 10.04. Everything went more or less smoothly, and most of the information you need is in the docs. Monit is configured via the
/etc/monit/monitrc file, and here's my example. The file is commented but broadly speaking my
monitrc is doing the following things:
- Checking on the health of the
proxy-node-appvia its special
- Exposes Monit's web-based front end on a specific port and controls access via a username and password.
- Sends email via Gmail's SMTP servers when there's an alert.
Once you've got Monit configured you can check the config file's syntax for errors like so:
sudo monit -t
And the start Monit like this:
sudo monit # Starts the Monit daemon sudo monit start all # Tells the Monit daemon to actually start monitoring things
If you've tweaked the config file and want Monit to reinitialise with the changes:
sudo monit reload
Once Monit is properly up and running it exposes a similar command line API to Upstart. The benefit of using Monit's over Upstart's is that you'll get more accurate and verbose status updates and alerts from Monit. What's more if you stop the app via Upstart (
sudo stop proxy-node-app) Monit will just go ahead and restart it soon after, but stopping via Monit will stop monitoring too. The Monit commnds look like this (and indeed we used one back in
sudo monit restart proxy-node-app sudo monit start proxy-node-app sudo monit stop proxy-node-app
Future thoughts, improvements?
More complex setup
For more complex apps that have more involved steps for initialisation (database setup etc.) we would need to move beyond a generic deployment script. Checking such a build script into version control along with the app code would probably be a good idea, as would making use of tools like Jake.
Rolling back failed deployments
By keeping a record of previous versions of the app (perhaps in directories with names like
proxy-node-app-v0.5) and symlinking the most recently deployed version to the live app folder it'd make it much easier to rollback a broken version of the app.
At the moment there is a simple 1:1 relationship between the developer's workstation environment and the production server. It would probably be quite easy to set up different branches in the remote Git repo, for example to represent
staging environments, and handle pushes to those branches differently in the
This deployment process has no concept of continuous integration or testing. While it probably wouldn't be hard to add some light-weight CI-like features I'm aware that it might be reinventing the wheel. Is this where Jenkins comes in?
Although I tried to keep this as simple as possible I'm aware that it involves a fair number of steps and fairly fiddly server configuration. It'd probably take at least half day to a day to get this process up and running from a cold start. Is this where tools such as Chef and Puppet could help?