22 Aug 2012 8:35 AM #1
Appearance Design Pattern Implementation Inconsistencies?
Appearance Design Pattern Implementation Inconsistencies?
I'm posting because I've been working with the GXT 3 Appearances and I'm seeing a variety of implementation strategies in regards to the resource and instance management within the various types.
With the com.sencha.gxt.core.client.resources.ThemeStyles.ThemeAppearance implementation there is a clean, and IMO, good implementation of the Singleton creational design pattern. The TemeStyles class has a private static ThemeAppearance instance that is initialized at the class level. Then the ThemeStyles.get() exposes the instance and the ThemeStyles.Styles is exposed from the instance. This is great because it allows for efficient resource management and referencing the ThemeAppearance resources from UIBinder etc...
A similar implementation of an Appearance is the com.sencha.gxt.core.client.resources.CommonStyles.CommonStylesAppearance. There is a difference in the singleton design pattern implementation though which seems a bit unfortunate. The CommonStyles.get() exposes only the CommonStyles.Styles type which doesn't allow for API access to the actual CommonStylesAppearance and related resources.
As I dig into the many other implementations of Appearances (and their associated resources and styles etc...) I don't see any real consistency with the implementation details of instance management. In many cases the different constructors simply instantiate via GWT.create() instance based versions of the resources with no consideration (I think) to sharing the resources in a flweight'ish/singleton creational pattern manner.
I'm posting because I'd like to find out if there are plans to clean up and standardize the creational and management patterns used within the GXT Implementations of the Appearances. If they were consistent, it would allow us to plan our use of existing Appearances and extensions of them to be cleaner and more consistent.
26 Aug 2012 4:16 PM #2
Thanks for your feedback on how we've elected to deal with appearances in GXT 3 - I'm going to try to address your main points by describing the primary design goals and non-goals, and implementation details that we've elected to follow. Certainly, we haven't been totally consistent, but I think that we've hit most of the ideas correctly.
First, some technical pieces - these'll come in handy later:
Singletons - these are usually considered helpful in cases where either a) state must be shared, or b) allocation is expensive. The former is true in a few places (CommonStyles, Mask, etc), but elsewhere it must be possible to create a component or cell with a specific implementation, different from the rest of the app.
GWT.create - this is slightly magic, but in the end is the same as 'new'. This breaks down to the following - look for <replace-with> or <generate-with> rules in the modules currently used - if any rule applies, invoke the provided type's default constructor. Otherwise, try to invoke the type passed into the GWT.create method with new. This means there is no flyweight/singleton thing going on there at all - new instances will be created (more on this in a bit) each time GWT.create is invoked...
GWT's optimizing compiler - two interesting and applicable things that happen here are to a) mark all possible types, methods, fields, and variables as final, and b) any final method that is found can now be rewritten as a static method. This has immediate two advantages - call sites can be rewritten as if there was no appearance instance, and the keyword 'this' isn't necessary within the method itself, both of which make for smaller code. Next, if the method itself doesn't actually use fields or polymorphic methods, the 'this' argument might be possible to drop. Especially where appearances are stateless or where there is only one implementation of a given appearance interface, this means that GWT.create effectively creates a singleton, except without the instance itself, so even cheaper than a singleton.
ClientBundle - this is the most expensive part of creating an appearance instance, since there is CSS that must be injected into the page. That being said, adding the styles to the page is only costly the first time it is called (see CssResource.ensureInjected), every subsequent time, it has no effect.
Okay, appearances themselves have several main goals:
* easily maintainable,
* efficient rendering,
* usable in both Components and Cells, and
* small compiled output
* external css and images (though a specific set of appearance implementations could achieve this)
* API to allow code external to the cell/component to modify the rendered internals
Any component which also has a similar cell always delegates to that cell so there are sure to render in the same way. To that end, we have two basic ways to select a new appearance from the defaults in GXT:
* use an alternate constructor, which accepts an appearance instance
* add a <replace-with> rule for a different appearance implementation or an <inherits> statement for a theme (module with many replace-with rules for specific appearances)
The first allows specific instances to be modified, and the second for the entire application to be rethemed in a single stroke. If singletons were used for some appearances like ContentPanel, it would only ever be possible to draw ContentPanels in one way in an application. As such, singletons are only used in specific cases where it only ever makes sense to have one instance - global details like CommonStyles, ThemeStyles, and necessarily consistent ones like DnD details or component masks.
Many of the provided appearance implementations also allow for optional constructor arguments, making it easier to easily subclass them and provide different styles or html structures. Of course it is still possible to provide a completely new implementation of the appearance interface, but this isn't required.
One last point here, regarding the first non-goal: Many developers prefer to have one giant css file for all possible 'overrides' - styles that should be replaced at runtime for application-specific reasons. GXT (via ClientBundle/CssResource) deliberately does not make this easy to do, as while it makes for quick and easy fixes, it increases the amount of CSS the browser must parse on page load and on any change to the dom, and almost always results in useless CSS rules, since the old rule now being overridden still exists somewhere. Using CssResource and the @Source annotation makes it clear to the compiler and the browser which styles will be used, and gives the compiler a chance to remove unused code.
Most appearances are used directly by a Cell, and then indirectly by a Component, so being able to render to a SafeHtmlBuilder is important. For the remaining cases, it is almost always more efficient (according to current benchmarks) to build up html strings an to pass the entire string at once into element.setInnerHtml. As a result of these guidelines, all appearances are expected to either produce a SafeHtml instance, or to know how to render themselves into an existing SafeHtmlBuilder.
As mentioned in the maintenance section, CSS that is never loaded is the fastest, so it is a goal to keep the active amount of CSS in the browser to a minimum. ClientBundle/CssResource has a number of details already in place to make this work, so the GXT provided appearances always make use of this.
Usable in both Components and Cells:
As mentioned above, the appearance APIs should in general be html-driven when rendering content, but may accept Element (or XElement) instances when working with already rendered content and events. This requirement has other interesting impacts on maintenance - appearances should focus on rendering and updating the ui, cells on managing DOM events, and components on high level logic (such as can be handled only when there is exactly one rendered instance of a cell).
Small compiled output:
By using specific classes and referencing those types either in <replace-with> rules or explicitly invoking them, we are making it clear to the compiler which appearances, html strings, and styles (css and images) will be required. By being this explicit, we can be fairly certain that the compiler will remove all unused appearance code - if you add a replace-with rule for all TriggerFieldAppearances, you can be certain that GXT's own implementations will never touch your users' browsers.
Specifically looking at the CommonStyles and ThemeStyles example you mentioned, there appears to be a good reason that the CommonStyles appearance instance isn't exposed - the only useful method it provides is access to the styles. The shim() method shouldn't have been made accessible in that way - it is only needed to make the CommonDefaultStyles css resource work correctly. ThemeStyles exposes the appearance itself to provide access to the one other method - the more icon, which was intended to be needed by more than widget so must be available in a shared manner (but is only used in one place, so doesn't make a ton of sense).
There are definitly some inconsistencies in appearances that need to be addressed, but I'd tenatively suggest that between ClientBundle and the compiler itself, most of your concerns are mitigated. Specific areas of improvement I see are mostly API-related:
* Failure to keep appearance instances as private final fields, and to expose a type-safe getter for subclasses to know what methods they are guaranteed to have access to http://www.sencha.com/forum/showthread.php?236527
* Failure to rely on those getters internally, to allow subclassed cells/components to override the appearance implementations at will
* Insufficient appearance examples (somewhat mitigated by 3.0.1's Gray theme, as well as additional theming examples in the explorer)
* Possible performance issues in rendering appearances and updating after events have taken place. This mostly have to do with making appearances even faster by assuming that external code has no business re-structuring the dom of a given widget - by making this assumption, we can more efficiently look up elements by index instead of by class name, and can enable our widgets and cells to draw and update even faster.
* Lack of general documentation on appearances in general - this is why I try to make details forum posts when these questions come up, but we had initially had hopes that the pattern would be more wide-spread in GWT itself (instead of only in TextButton), and we could draw on those same resources and documents.
So yes, we have plans to try to be more clean and consistent, but I think so far we've gotten things mostly right. I look forward to your thoughts on the ideas I've posted above, and to any future ideas you might have to make this feature in GXT 3 even clearer and easier to use.
27 Aug 2012 4:36 AM #3
That is an excellent reply! I'm left chewing on all the information you have provided with very little to question at this point.
Having all the information that you've provided in the back of my mind as I continue with my development efforts using gxt will be very helpful. It has cleared up a number of other issues that were in the back of my mind that I wasn't sure if I was misunderstanding or if they were implementation issues.
Thanks for taking the time to be so detailed in your reply, I'm sure this thread will prove useful to many others that are digging into the workings and goals of the Appearances.