The Success of the Zope Component Architecture
Prompted by recent brief negative pronunciations by Malthe on the Zope Component Architecture (ZCA), I thought I'd talk a bit about what I think about it. I'm not going to go into hermeneutics here of what Malthe might mean -- others attempts at exegesis exist in the comments to that blog entry already. Instead, I'll just talk about what I think makes the ZCA useful, and why it is successful. Finally I'll go into some reasons why people are frustrated by the ZCA.
What is the ZCA used for?
What is the ZCA used for? It's used to glue things to each other: glue event handlers to events, glue views to models, glue plugins into applications and libraries, and more abstractly, glue adapters to adaptees. In ZCA terms, providing such glue is termed providing configuration.
Why was the ZCA created? The Zope community had been building pluggable web applications for a long time and we noticed our components became overly complex and were hard to glue together and override. The ZCA is one answer to this problem.
What advantage does such gluing bring?
many, perhaps all, larger applications contain glue. The ZCA makes the glue explicit and uniform.
ZCA glue can be overridden explicitly.
you can extend existing systems by gluing in new things.
This way, hurry.resource doesn't need to know anything in particular about URL generation or requests; its plugins can take care of this. This allowed me to write and test hurry.resource without worrying too much about the larger Zope Toolkit framework, knowing I could plug it in later, and now the library becomes more useful for a broader group of people.
The ZCA doesn't just allow one to glue one thing to another, but also to override the glue in specific cases. A common example of overriding glue occurs with views. Zope Toolkit applications follow a view/model approach, where the view is looked up on the model dependent on its class (or interface). It happens frequently enough in an application that I want many models to share a particular view, but override one model with a more specialized view. This is much like the way inheritance works in plain Python: I implement a method on a base class shared completely by some subclasses, but for one subclass I'd like to override it.
What I described just now is overriding for particular subclasses. The ZCA also allows other ways of overriding based on the zope.configuration library, overriding one glue registration with another one. I myself find myself using this kind of override less frequently, but it's still very useful in the 1 percent of cases where other options would be very ugly.
The question is sometimes asked why not just modify code dynamically for overrides? Why not monkey patch it? (or "open the class", or whatever other terminology one would like to use).
Brandon Rhodes at PyCon 2008 gave a great presentation explaining why the ZCA approach can be superior to monkey-patching and some other approaches. The point he makes is that the ZCA is the most composable and maintainable approach of the alternatives (subclassing, mixins, monkey-patching, adaptation). I recommend everyone interested in this topic to read his slides.
Of course as with everything in programming, everything depends on the trade-offs. In many circumstances the ZCA is overkill. The ZCA can be misused. But I also maintain that in many circumstances the ZCA is very useful.
Component approaches have become quite popular with web frameworks. Many popular ones adopt elements of it, but often in a somewhat limited way. For instance, one very popular interface is the WSGI interface, and one popular form of adaptation of this WSGI interface is to use WSGI middleware and framework components. With WSGI we see that just one well-defined, consensus interface has become an amazing source of creativity and pluggability within web development. With the ZCA we are able to define more than one interface in our applications, and potentially create ecosystems of creativity around those. Not that this is easy, but at least we have a mechanism and method to do so.
The ZCA is successful
Why do I say the Zope Component Architecture (ZCA) is successful? It's successful as it's being used, by many people, for many years now. Of course you can say it's only used by that weird Zope community, and a bit by weird Twisted people as well perhaps. That's fine, but realize that the wider Zope community is big and is made of many parts: Plone, Silva, Zope 2, Zope Toolkit, Grok. I'll also count BFG as part of the wider Zope community.
The ZCA is successful for me. Without it, I'd have to invent something very much like it. It comes back in much code that I've written. I'm able to do all kinds of small, cool things in my applications and libraries on a daily basis, and I'm creating more reusable code as a result.
Frustration with the ZCA
The ZCA isn't perfect. Humans aren't perfect, either. It's not easy to create good reusable interfaces, or to build pluggable frameworks and applications. It's easy to overdo the pluggability. It's easy to create the wrong pluggability points. It can get overly complex and overwhelming.
Writing XML configuration files to glue things together can be cumbersome and lead to repeating yourself, though his last issue has been overcome for some time now by the Grok project and its reusable solution (grokcore.component) is available to everybody.
It's easy to let the pain of the mistakes in the use of the ZCA, and the pain of complex applications in which it is used in general, overshadow the many successes of the ZCA: in a large part thanks to to the ZCA's ability to glue things together people are able to use Zope Toolkit libraries within a Zope 2 or Plone context. Grok was able to remix the Zope Toolkit in part because we could easily modify and extend the framework. I think the web frameworks that use the ZCA offer pluggability and flexibility unrivaled by the competition.
We generally don't say that the problems and complexities in our large libraries and applications are due to the Python programming language. While Python isn't perfect, we tend to think overall it helps. I also tend to think that the ZCA helps.
Are there alternatives to the ZCA? Certainly: any library or application specific pluggability system is an alternative. Sometimes a specific approach is better than a generic one. It can be easier to understand it's smaller.
Usually I'd say it's the other way around. Each custom pluggability system is another one to learn, and has its own limitations. I've found multiple times that I can make a far more powerful pluggability system for a library in a shorter time if I build it on top of the ZCA.
There are also more general approaches where there's at least some overlap with the ZCA: setuptools entry points, WSGI middleware, and PEAK-Rules. To use any of them for pluggability and overriding one needs to have a notion of an interface one can implement and plug into, a notion that the ZCA makes explicit.
The ZCA is useful. The ZCA is powerful. The ZCA is successful. The ZCA is imperfect. When I've run into its imperfections I've helped build solutions on top of it, such as Grok's in-python gluing approach. Sometimes I've considered its use overkill and gone for some other approach. I think it's useful to step back and consider alternative approaches. But let's consider them in the light of the success of the ZCA as much as in the light of its flaws.