Ant in Anger:

Using Apache Ant in a Production Development System

Steve Loughran
Last updated 2005-03-16

Introduction

Apache Ant can be an invaluable tool in a team development process - or it can be yet another source of problems in that ongoing crises we call development . This document contains some strategies and tactics for making the most of Ant. It is moderately frivolous in places, and lacks almost any actual examples of Ant XML. The lack of examples is entirely deliberate - it keeps document maintenance costs down. Most of the concepts covered don't need the detail provided by XML representations, as it is the processes we are concerned about, not the syntax. Finally, please be aware that the comments here are only suggestions which need to be customised to meet your own needs, not strict rules about what should and should not be done.

Firstly, here are some assumptions about the projects which this document covers:

What that all means is that there is no time to spend getting things right, you don't have that tight control on how the rest of the team works and the development process is often more one of chaos minimisation than anything else. The role of Ant in such projects is to ensure that the build, test and deploy processes run smoothly, leaving you with all the other problems.

Core Practices

Clarify what you want Ant to do

Ant is not a silver bullet. It is just another rusty bullet in the armory of development tools available at your disposal. Its primary purpose is to accelerate the construction and deployment of Java projects. You could certainly extend Ant to do anything Java makes possible: it is easy to imagine writing an image processing task to help in web site deployment by shrinking and recompressing jpeg files, for example. But that would be pushing the boundary of what Ant is really intended to do - so should be considered with care.

Ant is also a great adjunct to an IDE; a way of doing all the housekeeping of deployment and for clean, automated builds. But a good modern IDE is a productivity tool in its own right - one you should continue to use. Ant just lets you give the teams somewhat more freedom in IDE choice - "you can use whatever you want in development, but Ant for the deployment builds" Now that many modern open source and commercial IDEs include Ant support (including jEdit, Forte, Eclipse and IDEA), developers can use a great IDE, with Ant providing a rigorous and portable build process integrated into the tool.

Define standard targets

When you have multiple sub projects, define a standard set of targets. Projects with a split between interface and implementation jar files could consider impl and intf targets - with separate debug-impl and debug-intf targets for the debug version. And of course, the ubiquitous clean target.

With standard target names, it is easy to build encompassing Ant build files which just hand off the work to the classes below using the ant task. For example. the clean target could be handed down to the intf and impl subdirectories from a parent directory

<target name="clean"  depends="clean-intf, clean-impl">
</target>

<target name="clean-intf" >
    <ant dir="intf" target="clean" />
</target>

<target name="clean-impl">
    <ant dir="impl" target="clean" />
</target>  
If you give targets a description tag, then calling ant -projecthelp will list all tasks with their description as 'main targets', and all tasks without a description as subtargets. Describing all your entry points is therefore very useful, even before a project becomes big and complicated.

Extend Ant through new tasks

If Ant does not do what you want, you can use the exec and java tasks or inline scripting to extend it. In a project with many build.xml files, you soon find that having a single central place for implementing the functionality keeps maintenance overhead down. Implementing task extensions through Java code seems extra effort at first, but gives extra benefits:- In a way, it is this decoupling of functionality, "the tasks", from the declaration of use, "the build file", that has helped Ant succeed. If you have to get something complex done in Make or an IDE, you have a hairy makefile that everyone is scared of, or an IDE configuration that is invariably very brittle. But an Ant task is reusable and shareable among all Ant users. Many of the core and optional tasks in Ant today, tasks you do or will come to depend on, were written by people trying to solve their own pressing problems.

Embrace Automated Testing

(alternatively "recriminate early, recriminate often")

Ant lets you call JUnit tasks, which unit test the code your team has written. Automated testing may seem like extra work at first, but JUnit makes writing unit tests so easy that you have almost no reason not to. Invest the time in learning how to use JUnit, write the test cases, and integrate them in a 'test' target from Ant so that your daily or hourly team build can have the tests applied automatically. One of the free to download chapters of Java Development with Ant shows you how to use JUnit from inside Ant.

Once you add a way to fetch code from the SCM system, either as an Ant task, in some shell script or batch file or via some continuous integration tool. the integration test code can be a pure Ant task run on any box dedicated to the task. This is ideal for verifying that the build and unit tests work on different targets from the usual development machines. For example, a Win95/Java1.1 combination could be used even though no developer would willingly use that configuration given the choice.

System tests are harder to automate than unit tests, but if you can write java code to stress large portions of the system - even if the code can not run as JUnit tasks - then the java task can be used to invoke them. It is best to specify that you want a new JVM for these tests, so that a significant crash does not break the full build. The Junit extensions such as HttpUnit for web pages, and Cactus for J2EE and servlet testing help to expand the testing framework. To test properly you will still need to invest a lot of effort in getting these to work with your project, and deriving great unit, system and regression tests - but your customers will love you for shipping software that works.

Learn to Use and love the add-ons to Ant

The Ant distribution is not the limit of the Ant universe, it is only the beginning. Look at the External Tools and Tasks page for an up to date list. Here are some of them that .

Cross Platform Ant

Ant is the best foundation for cross platform Java development and testing to date. But if you are not paying attention, it is possible to produce build files which only work on one platform - or indeed, one single workstation.

The common barriers to cross-platform Ant are the use of command line tools (exec tasks) which are not portable, path issues, and hard coding in the location of things.

Command Line apps: Exec / Apply

The trouble with external invocation is that not all functions are found cross platform, and those that are often have different names - DOS descendants often expect .exe or .bat at the end of files. That can be bad if you explicitly include the extension in the naming of the command (don't!), good when it lets you keep the unix and DOS versions of an executable in the same bin directory of the project without name clashes arising.

Both the command line invocation tasks let you specify which platform you want the code to run on, so you could write different tasks for each platform you are targeting. Alternatively, the platform differences could be handled inside some external code which Ant calls. This can be some compiled down java in a new task, or an external script file.

Cross platform paths

Unix paths use forward slashes between directories and a colon to split entries. Thus "/bin/java/lib/xerces.jar:/bin/java/lib/ant.jar" is a path in unix. In Windows the path must use semicolon separators, colons being used to specify disk drives, and backslash separators "c:\bin\java\lib\xerces.jar;c:\bin\java\lib\ant.jar".

This difference between platforms (indeed, the whole java classpath paradigm) can cause hours of fun.

Ant reduces path problems; but does not eliminate them entirely. You need to put in some effort too. The rules for handling path names are that 'DOS-like pathnames are handled', 'Unix like paths are handled'. Disk drives -'C:'- are handled on DOS-based boxes, but placing them in the build.xml file ruins all chances of portability. Relative file paths are much more portable. Semicolons work as path separators - a fact which is useful if your Ant invocation wrapper includes a list of jars as a defined property in the command line. In the build files you may find it better to build a classpath by listing individual files (using location= attributes), or by including a fileset of *.jar in the classpath definition.

There is also the PathConvert task which can put a fully resolved path into a property. Why do that? Because then you can use that path in other ways - such as pass it as a parameter to some application you are calling, or use the replace task to patch it into a localised shell script or batch file.

Note that DOS descended file systems are case insensitive (apart from the obscure aberration of the WinNT POSIX subsystem run against NTFS), and that Windows pretends that all file extensions with four or more letters are also three letter extensions (try DELETE *.jav in your java directories to see a disastrous example of this).

Ant's policy on case sensitivity is whatever the underlying file system implements, and its handling of file extensions is that *.jav does not find any .java files. The Java compiler is of course case sensitive - you can not have a class 'ExampleThree' implemented in "examplethree.java".

Some tasks only work on one platform - Chmod being a classic example. These tasks usually result in just a warning message on an unsupported platform - the rest of the target's tasks will still be called. Other tasks degrade their functionality on platforms or Java versions. In particular, any task which adjusts the timestamp of files can not do so properly on Java 1.1. Tasks which can do that - Get, Touch and Unjar/Unwar/Unzip for example, degrade their functionality on Java1.1, usually resorting to the current timestamp instead.

Finally, Perl makes a good place to wrap up Java invocations cross platform, rather than batch files. It is included in most Unix distributions, and is a simple download for Win32 platforms from ActiveState. A Perl file with .pl extension, the usual Unix path to perl on the line 1 comment and marked as executable can be run on Windows, OS/2 and Unix and hence called from Ant without issues. The perl code can be left to resolve its own platform issues. Don't forget to set the line endings of the file to the appropriate platform when you redistribute Perl code; fixCRLF can do that for you.

Team Development Processes

Even if each team member is allowed their choice of IDE/editor, or even OS, you need to set a baseline of functionality on each box. In particular, the JDKs and jars need to be in perfect sync. Ideally pick the latest stable Java/JDK version available on all developer/target systems and stick with it for a while. Consider assigning one person to be the contact point for all tools coming in - particularly open source tools when a new build is available on a nightly basis. Unless needed, these tools should only really be updated monthly, or when a formal release is made.

Another good tactic is to use a unified directory tree, and add on extra tools inside that tree. All references can be made relative to the tree. If team members are expected to add a directory in the project to their path, then command line tools can be included there - including those invoked by Ant exec tasks. Put everything under source code control and you have a one stop shop for getting a build/execute environment purely from CVS or your equivalent.

Deploying with Ant

One big difference between Ant and older tools such as Make is that the processes for deploying Java to remote sites are reasonably well evolved in Ant. That is because we all have to do it these days, so many people have put in the effort to make the tasks easier.

Ant can Jar, Tar or Zip files for deployment, while the War task extends the jar task for better servlet deployment. Jlink is a jar generation file which lets you merge multiple sub jars. This is ideal for a build process in which separate jars are generated by sub projects, yet the final output is a merged jar. Cab can be used on Win32 boxes to build a cab file which is useful if you still have to target IE deployment.

The ftp task lets you move stuff up to a server. Beware of putting the ftp password in the build file - a property file with tight access control is slightly better. The FixCRLF task is often a useful interim step if you need to to adjust the line endings of files. A WebDav task has long been discussed, which would provide a more secure upload to web servers, but it is still in the todo list. Rumour has it that there is such a task in the jakarta-slide libraries. With MacOS X, Linux and Windows XP all supporting WebDAV file systems, you may even be able to use copy to deploy though a firewall.

EJB deployment is aided by the ejb tasks, while the serverdeploy suite can deploy to multiple servers. The popularity of Ant has encouraged vendors to produce their own deployment tasks which they redistribute with their servers. For example, the Tomcat4.1 installation includes tasks to deploy, undeploy and reload web applications.

Finally, there are of course the fallbacks of just copying files to a destination using Copy and Copydir , or just sending them to a person or process using Mail or the attachment aware MimeMail. In one project our team even used Ant to build CD images through a build followed by a long set of Copy tasks, which worked surprisingly well, certainly easier than when we mailed them to the free email service on myrealbox.com, then pulled them down from the far end's web browser, which we were running over WinNT remote desktop connection, that being tunneled through SSH.

Directory Structures

How you structure your directory tree is very dependent upon the project. Here are some directory layout patterns which can be used as starting points. All the jakarta projects follow a roughly similar style, which makes it easy to navigate around one from one project to another, and easy to clean up when desired.

Simple Project

The project contains sub directories
bin common binaries, scripts - put this on the path.
build This is the tree for building; Ant creates it and can empty it in the 'clean' project.
dist Distribution outputs go in here; the directory is created in Ant and clean empties it out
doc Hand crafted documentation
lib Imported Java libraries go in to this directory
src source goes in under this tree in a hierarchy which matches the package names. The dependency rules of <javac> requires this.
The bin, lib, doc and src directories should be under source code control. Slight variations include an extra tree of content to be included in the distribution jars - inf files, images, etc. These can go under source too, with a metadata directory for web.xml and similar manifests, and a web folder for web content - JSP, html, images and so on. Keeping the content in this folder (or sub hierarchy) together makes it easier to test links before deployment. The actual production of a deployment image, such as a war file, can be left to the appropriate Ant task: there is no need to completely model your source tree upon the deployment hierarchy.

Javadoc output can be directed to a doc/ folder beneath build/, or to doc/javadoc.

Interface and Implementation split

If the interface is split from the implementation code then this can be supported with minor changes just by having a separate build path for the interface directory - or better still just in the jar construction: one jar for interface and one jar for implementation.

Loosely Coupled Sub Projects

In the loosely coupled approach multiple projects can have their own copy of the tree, with their own source code access rights. One difference to consider is only having one instance of the bin and lib directories across all projects. This is sometimes good - it helps keep copies of xerces.jar in sync, and sometimes bad - it can update foundational jar files before unit testing is complete.

To still have a single build across the sub projects, use parent build.xml files which call down into the sub projects.

This style works well if different teams have different code access/commitment rights. The risk is that by giving extra leeway to the sub projects, you can end up with incompatible source, libraries, build processes and just increase your workload and integration grief all round.

The only way to retain control over a fairly loosely integrated collection of projects is to have a fully automated build and test process which verifies that everything is still compatible. Sam Ruby runs one for all the apache java libraries and emails everyone when something breaks; your own project may be able to make use of Cruise Control for an automated, continuous, background build process.

Integrated sub projects

Tightly coupled projects have all the source in the same tree; different projects own different subdirectories. Build files can be moved down to those subdirectories (say src/com/iseran/core and src/com/iseran/extras), or kept at the top - with independent build files named core.xml and extras.xml.

This project style works well if everyone trusts each other and the sub projects are not too huge or complex. The risk is that a split to a more loosely coupled design will become a requirement as the projects progress - but by the time this is realised schedule pressure and intertwined build files make executing the split well nigh impossible. If that happens then just keep with it until there is the time to refactor the project directory structures.

Ant Update Policies

Once you start using Ant, you should have a policy on when and how the team updates their copies. A simple policy is "every official release after whatever high stress milestone has pushed all unimportant tasks (like sleep and seeing daylight) on the back burner". This insulates you from the changes and occasional instabilities that Ant goes through during development. Its main disadvantage is that it isolates you from the new tasks and features that Ant is constantly adding.

Often an update will require changes to the build.xml files. Most changes are intended to be backwards compatible, but sometimes an incompatible change turns out to be necessary. That is why doing the update in the lull after a big milestone is important. It is also why including ant.jar and related files in the CVS tree helps ensure that old versions of your software can be still be built.

The most aggressive strategy is to get a weekly or daily snapshot of the ant source, build it up and use it. This forces you to tweak the build.xml files more regularly, as new tasks and attributes can take while to stabilise. You really have to want the new features, enjoy gratuitous extra work or take pleasure in upsetting your colleagues to take this approach.

Once you start extending Ant with new tasks, it suddenly becomes much more tempting to pull down regular builds. The most recent Ant builds are invariably the best platform for writing your extensions, as you can take advantage of the regular enhancements to the foundational classes. It also prevents you from wasting time working on something which has already been done. A newly submitted task to do something complex such as talk to EJB engines, SOAP servers or just convert a text file to uppercase may be almost exactly what you need - so take it, enhance it and offer up the enhancements to the rest of the world. This is certainly better than starting work on your 'text case converter' task on Ant 0.8 in isolation, announcing its existence six months later and discovering that instead of adulation all you get are helpful pointers to the existing implementation. The final benefit of being involved with the process is that it makes it easier for your tasks to be added with the Ant CVS tree, bringing forward the date when Ant has taken on all the changes you needed to make to get your project to work. If that happens you can revert to an official Ant release, and get on with all the other crises.

You should also get on the dev mailing list , as it is where the other developers post their work, problems and experience. The volume can be quite high: 40+ messages a day, so consider routing it to an email address you don't use for much else. And don't make everyone on the team subscribe; it can be too much of a distraction.

Installing with Ant.

Because Ant can read environment variables, copy, unzip and delete files and make java and OS calls, it can be used for simple installation tasks. For example, an installer for tomcat could extract the environment variable TOMCAT_HOME, stop tomcat running, and copy a war file to TOMCAT_HOME/webapps. It could even start tomcat again, but the build wouldn't complete until tomcat exited, which is probably not what was wanted.

The advantage of using Ant is firstly that the same install targets can be used from your local build files (via an ant invocation of the install.xml file), and secondly that a basic install target is quite easy to write. The disadvantages of this approach are that the destination must have an up to date version of Ant correctly pre-installed, and Ant doesn't allow you to handle failures well - and a good installer is all about handling when things go wrong, from files being in use to jar versions being different. This means that Ant is not suited for shrink wrapped software, but it does work for deployment and installation to your local servers.

One major build project I was involved in had an Ant install build file for the bluestone application server, which would shutdown all four instances of the app server on a single machine, copy the new version of the war file (with datestamp and buildstamp) to an archive directory, clean up the current deployed version of the war and then install the new version. Because bluestone restarted JVMs on demand, this script was all you needed for web service deployment. On the systems behind the firewall, we upped the ante in the deployment process by using the ftp task to copy out the war and build files, then the telnet task to remotely invoke the build file. The result was we had automated recompile and redeploy to local servers from inside our IDE (Jedit) or the command line, which was simply invaluable. Imagine pressing a button on your IDE toolbar to build, unit test, deploy and then functional test your webapp.

One extra trick I added later was a junit test case to run through the install check list. With tests to verify access permissions on network drives, approximate clock synchronisation between servers, DNS functionality, ability to spawn executables and all the other trouble spots, the install script could automatically do a system health test during install time and report problems. [The same tests could also be invoked from a JMX MBean, but that's another story].

So, Ant is not a substitute for a real installer tool, except in the special case of servers you control, but in that context it does let you integrate remote installation with your build.

Tips and Tricks

get
The get task can fetch any URL, so be used to trigger remote server side code during the build process, from remote server restarts to sending SMS/pager messages to the developer cellphones.
i18n
Internationalisation is always trouble. Ant helps here with the native2ascii task which can escape out all non ascii characters into unicode. You can use this to write java files which include strings (and indeed comments) in your own non-ASCII language and then use native2ascii to convert to ascii prior to feeding through javac. The rest of i18n and l12n is left to you...
Use Property Files
Use external property files to keep per-user settings out the build files - especially passwords. Property files can also be used to dynamically set a number of properties based on the value of a single property, simply by dynamically generating the property filename from the source property. They can also be used as a source of constants across multiple build files.
Faster compiles with Jikes
The jikes compiler is usually much faster than javac, does dependency checking and has better error messages (usually). Get it. Then set build.compiler to "jikes" for it to be used in your build files. Doing this explicitly in your build files is a bit dubious as it requires the whole team (and sub projects) to be using jikes too - something you can only control in small, closed source projects. But if you set ANT_OPTS = -Dbuild.compiler=jikes in your environment, then all your builds on your system will use Jikes automatically, while others can choose their own compiler, or let ant choose whichever is appropriate for the current version of Java.
#include targets to simplify multi build.xml projects
You can import XML files into a build file using the XML parser itself. This lets a multi-project development program share code through reference, rather than cut and paste re-use. It also lets one build up a file of standard tasks which can be reused over time. Because the import mechanism is at a level below which Ant is aware, treat it as equivalent to the #include mechanism of the 'legacy' languages C and C++.

There are two inclusion mechanisms, an ugly one for all parsers and a clean one. The ugly method is the only one that was available on Ant1.5 and earlier:-

    <!DOCTYPE project [
      <!ENTITY propertiesAndPaths SYSTEM "propertiesAndPaths.xml">
      <!ENTITY taskdefs SYSTEM "taskdefs.xml">
    ]>  
    
        &propertiesAndPaths;
        &taskdefs;
The cleaner method in Ant1.6 is the <import> task that imports whole build files into other projects. The entity inclusion example could almost be replaced by two import statements:-
 <import file="propertiesAndPaths.xml">
 <import file="taskdefs.xml">
We say almost as top level declarations (properties and taskdefs) do not get inserted into the XML file exactly where the import statement goes, but added to the end of the file. This is because the import process takes place after the main build file is parsed, during execution, whereas XML entity expansion is handled during the parsing process.

The <import> task does powerful things, such as let you override targets, and use ant properties to name the location of the file to import. Consult the documentation for the specifics of these features.

Before you go overboard with using XML inclusion, note that the ant task lets you call any target in any other build file - with all your property settings propagating down to that target. So you can actually have a suite of utility targets - "deploy-to-stack-a", "email-to-team", "cleanup-installation" which can be called from any of your main build files, perhaps with subtly changed parameters. Indeed, after a couple of projects you may be able to create a re-usable core build file which contains the core targets of a basic Java development project - compile, debug, deploy - which project specific build files call with their own settings. If you can achieve this then you are definitely making your way up the software maturity ladder. With a bit of work you may progress from being a SEI CMM Level 0 organisation "Individual Heroics are not enough" to SEI CMM Level 1, "Projects only succeed due to individual heroics"

NB, ant copies all your properties unless the inheritall attribute is set to false. Before that attribute existed you had to carefully name all property definitions in all build files to prevent unintentional overwriting of the invoked property by that of the caller, now you just have to remember to set inheritall="false" on all uses of the <ant> task.

Implement complex Ant builds through XSL
XSLT can be used to dynamically generate build.xml files from a source xml file, with the xslt task controlling the transform. This is the current recommended strategy for creating complex build files dynamically. However, its use is still apparently quite rare - which means you will be on the bleeding edge of technology.
Change the invocation scripts
By writing your own invocation script - using the DOS, Unix or Perl script as a starting point - you can modify Ant's settings and behavior for an individual project. For example, you can use an alternate variable to ANT_HOME as the base, extend the classpath differently, or dynamically create a new command line property "project.interfaces" from all .jar files in an interfaces directory.

Having a custom invocation script which runs off a CVS controlled library tree under PROJECT_HOME also lets you control Ant versions across the team - developers can have other copies of Ant if they want, but the CVS tree always contains the jar set used to build your project.

You can also write wrapper scripts which invoke the existing Ant scripts. This is an easy way to extend them. The wrapper scripts can add extra definitions and name explicit targets, redefine ANT_HOME and generally make development easier. Note that "ant" in Windows is really "ant.bat", so should be invoked from another batch file with a "CALL ant" statement - otherwise it never returns to your wrapper.

Write all code so that it can be called from Ant
This seems a bit strange and idealistic, but what it means is that you should write all your java code as if it may be called as a library at some point in future. So do not place calls to System.exit() deep in the code - if you want to exit a few functions in, raise an exception instead and have main() deal with it.

Moving one step further, consider proving an Ant Task interface to the code as a secondary, primary or even sole interface to the functionality. Ant actually makes a great bootloader for Java apps as it handles classpath setup, and you can re-use all the built in tasks for preamble and postamble work. Some projects, such as XDoclet only run under Ant, because that is the right place to be.

Use the replace task to programmatic modify text files in your project.
Imagine your project has some source files - BAT files, ASPX pages(!), anything which needs to be statically customised at compile time for particular installations, such driven from some properties of the project such as JVM options, or the URL to direct errors too. The replace task can be used to modify files, substituting text and creating versions customised for that build or destination. Of course, per-destination customisation should be delayed until installation, but if you are using Ant for the remote installation that suddenly becomes feasible.
Use the mailing lists
There are two mailing lists related to Ant, user and developer. Ant user is where all questions related to using Ant should go. Installation, syntax, code samples, etc - post your questions there or search the archives for whether the query has been posted and answered before. Ant-developer is where Ant development takes place - so it is not the place to post things like "I get a compilation error when I build my project" or "how do I make a zip file". If you are actually extending Ant, on the other hand, it is the ideal place to ask questions about how to add new tasks, make changes to existing ones - and to post the results of your work, if you want them incorporated into the Ant source tree.

Putting it all together

What does an Ant build process look like in this world? Assuming a single directory structure for simplicity, the build file should contain a number of top level targets Sub projects "web", "bean-1", "bean-2" can be given their own build files - web.xml, bean-1.xml, bean-2.xml - with the same entry points. Extra toplevel tasks related to databases, web site images and the like should be considered if they are part of the process.

Debug/release switching can be handled with separate initialisation targets called before the compile tasks which define the appropriate properties. Antcall is the trick here, as it allows you to have two paths of property initialisation in a build file.

Internal targets should be used to structure the process

The switching between debug and release can be done by making init-release conditional on a property, such as release.build being set :-
<target name="init-release" if="release.build">
    <property name="build.debuglevel" value="lines,source"/>    
  </target>
You then have dependent targets, such as "compile", depend on this conditional target; there the "default" properties are set, and then the property is actually used. Because Ant properties are immutable, if the release target was executed its settings will override the default values:
<target name="compile" depends="init,init-release">
    <property name="build.debuglevel" value="lines,vars,source"/> 
    <echo>debug level=${build.debuglevel}</echo>
    <javac destdir="${build.classes.dir}"
       debug="true"
       debuglevel="${build.debuglevel}"
       includeAntRuntime="false"
       srcdir="src">
      <classpath refid="compile.classpath"/>
    </javac>
  </target>
As a result, we now have a build where the release mode only includes the filename and line debug information (useful for bug reports), while the development system included variables too.

It is useful to define a project name property which can be echoed in the init task. This lets you work out which Ant file is breaking in a multi file build.

What goes in to the internal Ant tasks depends on your own projects. One very important tactic is "keep path redefinition down through references" - you can reuse paths by giving them an ID and then referring to them via the "refid" attribute you should only need to define a shared classpath once in the file; filesets can be reused similarly.

Once you have set up the directory structures, and defined the Ant tasks it is time to start coding. An early priority must be to set up the automated test process, as that not only helps ensures that the code works, it verifies that the build process is working.

And that's it. The build file shouldn't need changing as new source files get added, only when you want to change the deliverables or part of the build process. At some point you may want to massively restructure the entire build process, restructuring projects and the like, but even then the build file you have should act as a foundation for a split build file process -just pull out the common properties into a properties file all build files read in, keep the target names unified and keep going with the project. Restructuring the source code control system is often much harder work.

The Limits of Ant

Before you start adopting Ant as the sole mechanism for the build process, you need to be aware of what it doesn't do.

It's not a scripting language

Ant lets you declare what you want done, with a bit of testing of the platform and class libraries first to enable some platform specific builds to take place. It does not let you specify how to handle things going wrong (a listener class can do that), or support complex conditional statements.

If your build needs to handle exceptions then look at the sound listener as a simple example of how to write your own listener class. Complex conditional statements can be handled by having something else do the tests and then build the appropriate Ant task. XSLT can be used for this.

It's not Make

Some of the features of make, specifically inference rules and dependency checking are not included in Ant. That's because they are "different" ways of doing a build. Make requires you to state dependencies and the build steps, Ant wants you to state tasks and the order between them, the tasks themselves can do dependency checking or not. A full java build using Jikes is so fast that dependency checking is relatively moot, while many of the other tasks (but not all), compare the timestamp of the source file with that of the destination file before acting.

It's not meant to be a nice language for humans

XML isn't a nice representation of information for humans. It's a reasonable representation for programs, and text editors and source code management systems can all handle it nicely. But a complex Ant file can get ugly because XML is a bit ugly, and a complex build is, well, complicated. Use XML comments so that the file you wrote last month still makes sense when you get back to it, and use Antidote to edit the files if you prefer it.

Big projects still get complicated fast

Large software projects create their own complexity, with inter-dependent libraries, long test cycles, hard deployment processes and a multitude of people each working on their own bit of the solution. That's even before the deadlines loom close, the integration problems become insurmountable, weekends become indistinguishable from weekdays in terms of workload and half the team stops talking to the other half. Ant may simplify the build and test process, and can eliminate the full time "makefile engineer" role, but that doesn't mean that someone can stop "owning the build". Being in charge of the build has to mean more than they type "ant all" on their system, it means they need to set the standards of what build tools to use, what the common targets, what property names and files should be and generally oversee the sub projects build processes. On a small project, you don't need to do that - but remember: small projects become big projects when you aren't looking. If you start off with a little bit of process, then you can scale it if needed. If you start with none, by the time you need it will be too late.

You still need all the other foundational bits of a software project

If you don't have an source code management system, you are going to end up hosed. If you don't have everything under SCM, including web pages, dependent jars, installation files, you are still going to end up hosed, it's just a question of when it's going to happen. CVS is effectively free and works well with Ant, but Sourcesafe, Perforce, Clearcase and StarTeam also have Ant tasks. These tasks let you have auto-incrementing build counters, and automated file update processes.

You also need some kind of change control process, to resist uncontrolled feature creep. Bugzilla is a simple and low cost tool for this, using Ant and a continuous test process enables a rapid evolution of code to adapt to those changes which are inevitable.

End piece

Software development is meant to be fun. Being in the maelstrom of a tight project with the stress of integration and trying to code everything up for an insane deadline can be fun - it is certainly exhilarating. Adding a bit of automation to the process may make things less chaotic, and bit less entertaining, but it is a start to putting you in control of your development process. You can still have fun, you should just have less to worry about, a shorter build/test/deploy cycle and more time to spend on feature creep or important things like skiing. So get out there and have fun!

Further Reading

About the Author

Steve Loughran is a research scientist at a corporate R&D lab, currently on a sabbatical building production web services against implausible deadlines for the fun of it. He is also a committer on Apache Ant and Apache Axis, and co-author of Java Development with Ant. He thinks that if you liked this document you'll love that book because it doesn't just explain Ant, it goes into processes, deployment and best practices and other corners of stuff that really make Ant useful. (It would have been easier to just rehash the manual, but that wouldn't have been so useful or as much fun).

For questions related to this document, use the Ant mailing list.