Facets allow you to add versioning to a system, but they are merely a mechanism, not a solution. You still have to make a decision as to how to version something. For example, at some point, you may want to deprecate a previous version’s behavior; at that point, you must make a decision how to handle requests for the deprecated version. For behavioral changes, you have to decide whether to use separate interfaces or use facets with the same interface. And, of course, you must have compatibility rules to determine what should happen if, for example, a version 1 object is passed to an operation that implements version 2 behavior. In other words, facets cannot do your thinking for you and are no panacea for the versioning problem.
The biggest advantage of facets is also the biggest drawback: facets delay the decision about the types that are used and their behavior until run time. While this provides a lot of flexibility, it is significantly less type safe than having explicit types that can be statically checked at compile time: if you have a problem relating to incorrect facet selection, the problem will be visible only at run time and, moreover, will be visible only if you actually execute the code that contains the problem, and execute it with just the right data.
Another danger of facets is to abuse them. As an extreme example, here is an interface that provides an arbitrary collection of objects of arbitrary type:
Even though this interface is empty, it can provide access to an unlimited number of objects of arbitrary type in the form of facets. While this example is extreme, it illustrates the design tension that is created by facets: you must decide, for a given versioning problem, how and at what point of the type hierarchy to split off a facet that deals with the changed functionality. The temptation may be to "simply add another facet" and be done with it. However, if you do that, your objects are in danger of being nothing more than loose conglomerates of facets without rhyme or reason, and with little visibility of their relationships in the type system.
In object modeling terms, the relationship among facets is weaker than an is‑a relationship (because facets are often not type compatible among each other). On the other hand, the relationship among facets is stronger than a
has‑a relationship (because all facets of an Ice object share the same object identity).
It is probably best to treat the relationship of a facet to its Ice object with the same respect as an inheritance relationship: if you were omniscient and could have designed your system for all current and future versions simultaneously, many of the operations that end up on separate facets would probably have been in the same interface instead. In other words, adding a facet to an Ice object most often implies that the facet has an
is‑partly‑a relationship with its Ice object. In particular, if you think about the life cycle of an Ice object and find that, when an Ice object is deleted, all its facets must be deleted, this is a strong indication of a correct design. On the other hand, if you find that, at various times during an Ice object’s life cycle, it has a varying number of facets of varying types, that is a good indication that you are using facets incorrectly.
Ultimately, the decision comes down to deciding whether the trade-off of static type safety versus dynamic type safety is worth the convenience and backward compatibility. The answer depends strongly on the design of your system and individual requirements, so we can only give broad advice here. Finally, there will be a point where no amount of facet trickery will get past the point when "yet one more version will be the straw that breaks the camel’s back." At that point, it is time to stop supporting older versions and to redesign the system.
