Service Implementation Reloading
Tapestry's ability to live reload page and component classes is a huge boon to development productivity, but what about reloading services?
On the one hand, a good application design keeps the page and component classes "thin" and moves logic into the services layer, for easier reuse across pages. On the other hand, moving logic into services is less agile if those services don't auto-reload the way pages do.
Starting in release 5.2, service implementations are live-reloaded, too. There are several restrictions on this, and a couple of leaky abstractions, but on the whole it's quite serviceable.
As of release 5.2, you can change your service implementation, and Tapestry picks up the change immediately. A service can even change its dependencies when being reloaded ... but it can't change its interface.
- Reloading only works for services, and only for services with the default service scope (i.e., global singletons). Live reloading does not apply to modules, service interfaces, contributions, or anything but just the service implementation.
- Reloading is limited to services that can be proxied: services with an interface and an implementation of that interface.
- Reloading requires that the underlying class files be simple, local, filesystem files (not, for example, files inside a JAR).
- If a service has internal state of any kind, that state is lost when the class is reloaded and the service re-instantiated. However, if a service has a configuration, the configuration will be reconstructed and injected into the service.
- Services are decorated only once, so any decorations or advice applies to the initially loaded version of the class, and will not be recalculated when the class changes.
Class Loader Issues
Tapestry creates a new class loader for each service implementation. When the underlying .class file changes, the class loader is discarded along with the instance, and a new class loader is created.
The class loader only loads the service implementation class, and any inner classes for the service implementation. All other classes are loaded by the standard class loader for the application.
Because of how class loaders work, the class will no longer be able to access package private classes and members of other classes in the same package. You may see some odd IllegalAccessErrors and need to change the visibility of package-private classes to be public.
The JVM should be able to eventually garbage collect the class loader. However, if the class publishes itself to some other service (for example, adding itself as a listener to an event published by some other service), then the instance and the garbage collector will be leaked.
Update checks are normally driven by tapestry-core, which periodically checks for changed templates, message catalogs, and component classes. Checks for changed service implementation classes occur at the same time.
In an application that is not driven by the web tier, you will need to periodically invoke the
fireCheckForUpdates() method of the UpdateListenerHub service (which was moved from tapestry-core to tapestry-ioc for this purpose).