Here are some recommendations and best practices we have gathered throughout our experience and consultancies with our customers.
In Ivy world, module descriptors are ivy files, which are basically simple xml files describing both what the module produce as artifact and its dependencies.
It is a good practice to write or download module descriptors for all the modules involved in your development, even for your third party dependencies, and even if they don't provide themselves such module descriptors.
First it will seem like an extra work, and require time. But when you will have several modules using the same third party library, and than you will only need to add one line to your ivy file to get this library and all its own dependencies that you really need (if you have good module descriptors in your repository, especially with the use of module configurations). It will also be very helpful when you want to upgrade a dependency. One single change in your module ivy file and you will get the updated version with its updated (or not) dependencies.
Therefore we recommend to add ivy files for all the modules in your repository, you can even enforce this rule by setting the allownomd attribute to false on your resolvers. Hence you shouldn't need to use the dependency artifact inclusion/exclusion/specification feature of Ivy, which should only be used in very specific cases.
This is usually not a valid recommendation for open source projects, but for the enterprise world we strongly suggest to avoid relying on a public repository like maven ibiblio or ivyrep. Why? Well, there are a couple of reasons:
The main problem with this kind of public repositories is that you don't have control over the repository. This means that if a module descriptor is broken you cannot easily fixed it. Sure you can use a chain between a shared repository and the public one and put your fixed module descriptor in the shared repository so that it hide the one on the public repository, but this makes repository browsing and maintenance cumbersome.
Even more problematic is the possible updates of the repository. We know that versions published in such repositories should be stable and not be updated, but we also frequently see that a module descriptor is buggy, or an artifact corrupted. We even see sometimes a new version published with the same name as the preceding one because the previous one was simply badly packaged. This can occur even to the best, it occured to us with Ivy 1.2 :-) But then we decided to publish the new version with a different name, 1.2a. But if the repository manager allow such updates, this means that what worked before can break. It can thus break your build reproducibility.
Ibiblio maven repository is not particularly well known for its reliability (we often experience major slow down or even complete break of the site), and ivyrep is only supported by a small company (yes we are only a small company!). So slow down and site hang occurs also. And if the repository you rely on is down, this can cause major slow down in your development or release process.
a public repository usually contains much more than what you actually need (except maybe ivyrep which certainly features much less than what you need :-)). Is it a problem? We think so. We think that in an enterprise environment the libraries you use should step through some kind of validation process before being used in every projects of your company. And what better way to do so? Setup an enterprise repository with only the libraries you actually want to use. This will not only ensure a better quality of your application dependencies, but help to have the same versions everywhere, and even help when declaring your module dependencies, if you use a tool like IvyDE, the code completion will only show relevant information about your repository, with only the libraries you actually want to see.
Note that it's not because you use an enterprise repository that you have to build it entirely by hand. Ivy features an install task which can be used to install modules from a repository to another one, so it can be used to selectively install modules from a public repository to your enterprise repository, where you will then be able to ensure control, reliability and accuracy.
Ivy is very flexible and can accomodate a lot of existing repositories, using the concept of patterns. But if your repository doesn't exist yet, we strongly recommend to always use the organisation and the module name in your pattern, even for private repository where you put only your own modules (which all the same organisation). Why? Because Ivy listing feature rely on the token it can find in the pattern. If you have no organisation token in your pattern, Ivy won't be able to list the (only?) organisation in your repository. And this can be a problem for code completion in IvyDE, for example, but also for repository wide tasks like install or rereport.
If you create a public repository, provide an url to corresponding ivyconf.xml. It's pretty easy to do, and if someone want to leverage your repository, he will just have to call configure with the url of your ivyconf.xml, or include it in its own configuration file, which makes it really easy to combine several public repositories.
Very often especially when working in a team or with several modules, you will need to rely on intermediate, non finalized versions of your modules. These versions are what we call integration versions, because their main objective is to be integrated with other modules to make and test an application or a framework.
If you follow the continuous integration paradigm across modules, these integration versions can be produced by a continuous integration server, very frequently.
So, how can you deal with these, possibly numerous, integration versions?
There are basically two ways to deal with them, both ways being supported by Ivy:
the idea is pretty simple, each time you publish a new integration of your module you give the same name to the version (in maven world this is for example 1.0-SNAPSHOT). The dependency manager should then be aware that this version is special because it changes over time, so that it does not trust its local cache if it already has the version, but check the date of the version on the repository and see if it hass changed. In Ivy this is supported using the changing attribute on a dependency or by configuring the changing pattern to use for all your modules.
in this case you use either a build number or a timestamp to publish each new integration version with a new version name. Then you can use one of the numerous ways in Ivy to express a version constraint. Usually selecting the very latest one (using 'latest.integration' as version constraint) is enough.
So, which way is the best? As often, it depends on your context, and if one of the two was really bad it wouldn't be supported in Ivy :-)
But usually we recommend to use the second one, because using a new version each time you publish a new version better fits the version identity paradigm, and can make all your builds reproducible, even integration one. And this is interesting because it enables, with some work in your build system, to introduce a mechanism to promote an integration build to a more stable status, like a milestone or a release.
Imagine you have a customer which comes on a monday morning and asks your latest version of your software, for testing or demonstration purpose. Obviously he needs it for the afternoon :-) Now if you have a continuous integration process and a good tracking of your changes and your artifacts, it may occur that you are actually able to fulfill his request without needing the use of a dolorean to give you some more time :-) But it may occur also that your latest version stable enough to be used for the purpose of the customer was actually built a few days ago, because the very latest just break a feature or introduce a new one you don't want to deliver. In this case, you can deliver this 'stable' integration build if you want, but be sure that a few days, or weeks, or even months later, the customer will ask for a bug fix on this demo only version. Why? Because it's a customer, and we all know how they are :-)
So, with a build promotion feature of any build in your repository, the solution would be pretty easy: when the customer ask for the version, you not only deliver the integration build, but you also promote it to a milestone status, for example. this promotion indicates that you should keep track of this version in a long period, to be able to come back to it and create a branch if needed.
Unfortunately Ivy does not by its own allow to have such reproducible builds out of the box, simply because Ivy is a dependency manager, not a build tool. But if you publish only versions with a distinct name and use Ivy features like versions constraint replacement during the publication or recursive delivery of modules, it can really help.
On the other hand, the main drawback of this solution is that it can produce a lot of intermediate versions, and you will have to run some cleaning scripts in your repository unless your company name starts with a G and ends with oogle :-)
With Ivy 1.4 you can resolve a dependency without even writing an ivy file. This pratice is called inlining. But what is it good for, and when should it be avoided?
Putting ivy dependencies in a separate file has the following advantages:
if your dependencies may change more often than your build, it's a good idea to separate the two, to isolate the two concepts: describing how to build / describing your project dependencies
if you describe dependencies of a module which can itself be reused, you will ant to publish it to a repository. In this case the publication is only possible if you have a separate ivy file
inline dependencies can only be used to express one dependency and only one. An ivy file can be used to express much more complex dependencies
On the other hand, using inline dependencies is very useful when:
If you already build your application and its modules using Ivy, it is really easy to leverage your ivy repository to download your application and all its dependencies on the local filesystem, ready to be executed. If you also put your configuration files as artifacts in your repository (maybee packaged as a zip), the whole installation process can rely on ivy, easing the automatic installation of any version of your application available in your repository!
Build and dependency management is often considered with a too low level priority in the software development world. We often see build management implemented by developers when they have time. Even if this may seem like a time and money saving in the short term, it often turns out to be a very bad choice in the long term. Building software is not a simple task, when you want to ensure automatic, tested, fully reproducible builds, releases and installations. On the other hand, once a good build system fitting your very specific needs is setup, it can then only rely on a few people with a good understanding of what is going on, with a constant quality ensured.
Therefore hiring a build and dependency expert to analyse and improve your build and release system is most of the time a very good choice (especially if you choose the right one :-))
These best practices are the reflect of our own experience, but we do not pretend to own the unique truth about dependency management or even Ivy use.
So feel free to comment on this page to add your own experience feedback, suggestions or opinion.