December 2, 2011

Day 2 - Strategies for Java Deployment

This article was written by Kris Buytaert (www.krisbuytaert.be/blog).

After years of working in Java-based environments, there are a number of things that I like to implement together with the teams I`m working with - the application doesn't matter much, whether it's plain java, Tomcat, JBoss, etc, these deployment strategies will help your ops and dev teams build more managable services.

Packaging

The first step is to have the native operating system packages as build artifacts rolling out of your continuous integration server - No .ear, .war or .jar files: I want to have rpms or debs. With things like fpm or the maven rpm plugin this should not be an extra hassle, and the advantages you get from doing this are priceless.

What advantages? Most native package systems support dependency resolution, file verification, and upgrades (or downgrades). These are things you would have to implement yourself or cobble together from multiple tools. As a bonus, your fellow sysadmins are likely already comfortable with the native package tool used on your systems, so why not do it?

Proxied, not running as root

Shaken, not stirred

Just like any other daemon, for security reasons, I prefer to run run Tomcat or JBoss as its own user, rather than as root. In most cases, however, only root can bind to ports below 1024, so you need to put a proxy in front. This is a convenient requirement because proxying (with something like Apache) can be used to terminate SSL connections, give improved logging (access logs, etc), and provides the ability to run multiple java application server instances on the same infrastructure.

Service Management

Lots of Java application servers have a semi functional shell script that allows you to start the service. Often, these services don't daemonize in a clean way, so that's why I prefer to use the Java Service wrapper from Tanuki to manage most Java based services. With a small config file, you get a clean way to stop and start java as a service and even the possibility to add some more monitoring to it.

However, there are some problems the Java Service wrapper leaves unsolved. For example, after launching the service, the wrapper can return back with a successful exit code while your service is not ready yet. The application server might be ready, but your applications themselves are still starting up. If you are monitoring these applications (e.g for High Availability), you really only want to treat them as 'active' when the application is ready, so you don't want your wrapper script to return, "OK," before the application has been deployed and ready. Otherwise, you end up with false positives or nodes that failover before the application has ever started. It's pretty easy to create a ping-pong service flapping scenario on a cluster this way.

One application per host

I prefer to deploy one application per host even though you can easily deploy multiple applications within a single Java VM. With one-per-host, management becomes much easier. Given the availability and popularity of good virtualization, the overhead of launching multiple Linux VM's for different applications is so low that there are more benefits than disadvantages.

Configuration

What about configuration of the application? Where should remote API urls, database settings, and other tunables go? A good approach is to create a standard location for all your applications, like /etc/$vendor/app/, where you place the appropriate configuration files. Volatile application configuration must be outside the artifact that comes out the build (.ear , .jar, .war, .rpm). The content of these files should be managed by a configuration management tool such as puppet, chef, or cfengine. The developers should be given a basic training so they can provide the systems team with the appropriate configuration templates.

Logs

Logs are pretty important too, and very easy to neglect. There are plenty of alternative tools around to log from a Java application: Log4j, Logback, etc .. Use them and make sure that they are configured to log to syslog, then they can be collected centrally and parsed by tools much easier than if they were spread all over the filesystem.

Monitoring

You also want your application to have some ways to monitor it besides just checking if it is running - it is usually insufficient to simply check if a tcp server is listening. A nice solution is to have a simple plain text page with a list of critical services and whether they are OK or not (true/false), for example:

someService: true
otherService: false

This benefits humans as well as machines. Tools like mon, heartbeat or loadbalancers can just grep for "false" in the file. If the file contains false, it reports a failure and fails over. This page should live on a standard location for all your applications, maybe a pattern like this http://host / servicename/health.html and an example "http://10.0.129.10:8080/mrs-controller/health.html". The page should be accessible as soon as the app is deployed.

This true/false health report should not be a static HTML file; it should be a dynamically generated page. Text means that you can also use curl, wget, or any command-line tool or browser to check the status of your service.

The 'health.html' page should report honestly about health, executing any code necessary to compute 'health' before yielding a result. For example, if your app is a simple calculator, it should verify health by doing tests internally like adding up some numbers before sharing 'myCalculator:true' in the health report.

The 'health.html' page should report honestly about health, executing any code necessary to compute 'health' before yielding a result. For example, if your app is a simple calculator, then before reporting health it should put two and two together and get four.

This kind of approach could also be used to provide you with metrics you can't learn from the JVM, such as number of concurrent users or other valid application metadata for measurement and trending purposes.

Conclusion

If you can't convince your developers, then maybe more data can help: Check out Martin Jackson's (presentation on java deployments)[http://www.slideshare.net/actionjackx/automated-java-deployments-with-rpm]

With good strategies in packaging, deployment, logging, and monitoring, you are in a good position to have an easily manageable, reproducible, and scalable environment. You'll give your developers the opportunity to focus on writing the application, they can use the same setup on their local development boxes (e.g. by using vagrant) as you are using on production.

By the way, I will be giving a talk titled DevOps: The past and future are here. It's just not evenly distributed (yet). at this year's LISA in Boston!

Further Reading

7 comments :

Toomas Römer said...

For these native packages, what should they contain? So lets say there is a myApp.war. Should the DEB or RPM manage the deployment of myApp.war (which will be quite difficult because the app server / servlet container can be from any vendor and installed differently). Or should I just unpack it somewhere?

asq said...

i think application server itself should be packaged, ie. to get benefit of not replacing its configuration files during upgrade (properly market %config in package). apps themselves are deployed as a process, usually driven by RM (ie. incremental changes) and it depends on how they're accustomized to CM software (ie. puppet).

Toomas Römer said...

Ok, gotcha, this solves the problem I was thinking about.

But what about application configuration now? Lets say they are at /etc/%vendor%/app. Should the app read them from a hardcoded location or there is a good solution here too?

Anonymous said...

Try Rex( http://rexify.org/ ).

* Install Tomcat And Deploy Web Applications With Rex
http://www.howtoforge.com/install-tomcat-and-deploy-web-applications-with-rex

* Continuous Deployment With Jenkins And Rex
http://www.howtoforge.com/continuous-deployment-with-jenkins-and-rex

Brian Dunbar said...

A good approach is to create a standard location for all your applications, like /etc/$vendor/app/, where you place the appropriate configuration files.

We use a tweak to this approach where /etc/$vendor is [1] a symbolic link to a NFS mounted share on NAS:

/nfs/nas_host_name/apps/vendor/conf

And not just config files but all apps are NFS mounted from NAS.


We do this for two reasons.

- Easier to edit those config files and app stuff: any host/user can do it.

- Log files are accessible from any host - as long as one knows the actual path.

- Backup is handled by snapshot on the NAS.

- DR is easier: Boot the new server, create the symlink.


[1] Obviously our location is different.

Brian Dunbar said...

Often, these services don't daemonize in a clean way, so that's why I prefer to use the Java Service wrapper from Tanuki to manage most Java based services.

Question.

We are moving from Solaris to Linux (OEL) at work. Over the next three months we're going to stand up about 12 new servers for applications. There will be more: the Plan is to keep Solaris only for Oracle DB, everything else will be Linux.

Solaris comes with SMF: which lets one manage any process as a service.

Does Linux come with a comparable facility? Or am I doomed to use Java Service Wrapper for my java apps, something else for my other apps and so on.

Jordan Sissel said...

Brian - Some options for you here include: daemontools, runit, supervisord, monit, god, mon, upstart, and systemd. Some of these are better designed than others, you'll have to pick one that works for you :)

Sponsored by Puppet Labs