Re: [xwiki-devs] [Proposal] Choosing the default component implementation

Hi Sergiu, Note that below I'm going to play the devil's advocate since I think it's important we really think hard before changing anything and verify if our current implementation is not enough. See below. On Apr 5, 2012, at 8:44 PM, Sergiu Dumitriu wrote: > On 04/05/2012 12:51 PM, Anca Luca wrote: >> >> On 04/05/2012 06:42 PM, Vincent Massol wrote: >>> >>> Hi Sergiu, >>> >>> On Apr 5, 2012, at 5:55 PM, Sergiu Dumitriu wrote: >>> >>>> Hi devs, >>>> >>>> Currently, requesting a component instance without a hint will look >>>> for the implementation that uses the "default" hint, which makes it >>>> difficult to change the implementation in an XWiki instance. Sure, it >>>> is easy as long as all the implementations use the "default" hint, >>>> but choosing the default between alternative implementations that >>>> should all still be usable by themselves is not possible. >>>> >>>> Also, "default" is not really a good hint, since it describes the >>>> state of the implementation, not the technology, the aspect that >>>> makes it different from the others. It would be better to name each >>>> implementation with a proper hint. >>>> >>>> I propose to define a mapping that can specify which hint is the >>>> default for a component. In a text file, >>>> META-INF/component-defaults.txt, we'll keep >>>> componentinterface=defaulthint mappings. For example: >>>> >>>> com.xpn.xwiki.store.XWikiStoreInterface=hibernate >>>> com.xpn.xwiki.store.migration.DataMigrationManager=hibernate >>>> >>>> And then, when we lookup the current storage implementation, we don't >>>> need to check what is the configured hint in xwiki.cfg (or >>>> xwiki.properties), we can just request the default implementation. >>>> >>>> If there's no mapping for a component, we'll continue to use the >>>> "default" hint. >>>> >>>> I'm not sure where exactly to keep such files. We bundle a >>>> components.txt file in each jar containing component implementations. >>>> We could do the same for the components we consider the platform >>>> defaults, and allow overrides in the >>>> WEB-INF/classes/META-INF/component-defaults.txt file. Still, this >>>> means that whenever platform defaults change, we need to keep another >>>> special section in the release notes, to let users know about these >>>> changes, so that they can manually revert to the old default if they >>>> need to. >>>> >>>> In the future we could change existing components to give proper >>>> hints instead of "default", where such a change is applicable. >>>> >>>> Another idea is to not use "default" at all, and instead go for a >>>> generic "xwiki", "xe", "xwiki-platform" or something like that >>>> whenever there's just one implementation for a component and we can't >>>> find another hint to describe it. >>>> >>>> WDYT? >>> >>> This is not really how it's been designed ATM. Whenever you wish to >>> use a different implementation of a component you use a component >>> implementation with the same role and same hint. You then make it >>> available in your classpath. (Of course you can also do this at >>> runtime simply by registering a new implementation over the old one). >>> >>> To decide which implementation is used you use a priority order, as >>> described on: >>> >>> http://extensions.xwiki.org/xwiki/bin/view/Extension/Component+Module#HOver… >>> >>> >>> I'd be curious to know your exact use case and understand why the >>> current mechanism doesn't work for it. > > "...choosing the default between alternative implementations that should > all **still be usable by themselves** is not possible" > > The overrides mechanism allows to change which component will be returned > for the "default" hint, but all the others will be invisible. > >> One usecase I see is that you have multiple implementations and you want >> to change the default one for a specific running instance of xwiki. >> >> Overwrite mechanism only allows you to say which impl should be used >> from the _components with the same hint_. However, you cannot change the >> hint of a component at configuration time, so if you have a standard >> distr of xwiki and you want to use ldap authentication, let's say (if >> only auth was impl with components), unless you do some java to add the >> default hint to the ldap implementation and then to specify that this >> one has priority over all the default ones, I don't see how you can >> re-wire the default. > > Exactly. For most of the "services" the XWiki platform currently has > (storage, cache), we don't have a "default" implementation, but we rely on a > kind of factory to lookup the configured default. That is an actual factory > class in the case of the cache service, but just more code in the old > XWiki.java class for the storage initialization. Yep that's how it works. > A standard way of selecting the default means that we'll need less > factories, and less code is always a good thing. Yes but it has more limitation to what we have since with a proper Factory you can imagine all kind of logic to decide which implementation to use (hour of the day, whether the user is a premium user or not, etc). BTW we do support Providers and the goal of the Provider is to be a factory for a given Role. So from now on, there should be no need to implement a Factory proper. Implementing a Provider is the new best practice for this. > Now, suppose one of the older components that had only one > implementation, "default", gets alternative implementations, and we want to > be able to allow more than one to be active in a wiki, and let the > administrator decide which one should be considered the default. How can we > approach this? The only way is indeed to have multiple hints, but last time > I checked this resulted in more than one instance, even for @Singleton > implementations. Correct, ATM we support only one hint per implementation.

> Another approach is to deprecate the direct dependency declaration and > instead introduce a factory that is responsible for selecting the default, > but this breaks backwards compatibility. Not completely true. Imagine you have (Role, "default") as your current component and you wish to be able to choose various implementation from now on. What you could do is this: * Write a new component, a Factory, that overrides the current component, i.e. by using the same (Role, "default"). This factory will then delegate to whatever other implementation it wishes.

From the user of (Role, "default")'s POV he won't see a single change. > The "I don't care, just give me the default" strategy works as long as > the component implementation is self-contained and doesn't involve data > communication. Let's take an example, PDF export. > > Currently the PDF export is only possible via FOP. If we were to convert > the current interface + implementation class into a proper component, we'd > name it "default", since it's the only one and who needs a factory for only > one implementation. Yes but we could also be thorough and instead implement a Provider<PDFExporter> instead. > People use it and they're happy because it just works. But we might soon > add support for export via an office server. Suppose we want either FOP or > Office to be usable as the default PDF export implementation. And suppose > we'll want to keep both types of export active, so that we can use either > one as the implementation for the default export of documents, the FOP > implementation for exporting some kinds of documents like scientific > articles, and the office connector for generating PDFs for presentations. > Using the overrides mechanism we can only have one active at the same time, > unless we introduce yet another factory which we can bypass using manual > component lookup. Based on your use case you'll need a UI to ask the user what he wants to export, i.e. either a "scientific article" or a "presentation" and from his choice you'll pick the right implementation. You could also try to guess that dynamically by looking at what is being exported but that'll require a Factory to hold that logic. I know what you mean though: for some reason you don't want that to be dynamic and you wish it to be static. There's a problem with dynamicity though. Imagine an extension that wants to replace an implementation. With your proposal the best practice would be to introduce a new Hint since the "default" is chosen statically at configuration time. So that wouldn't work. After you install extensions you'd need to stop the wiki, change the binding to the new implementation from the extension and restart it. Of course you'd need to read the documentation of the extension to know you have to use it in replacement. Basically it would mean that extensions cannot override behaviors. They would just be able to add new components (like new Macros) but not modify behaviors at runtime.

> If at some point we decide that the office implementation works better, > we might consider changing the default implementation for the pdf component. > This means that we'll change the hint of the FOP implementation from > "default" to "fop", change the hint of the Office implementation from > "office" to "default", and our new version of XE works great if people read > the installation guide and properly configure the office connector. In practice we would have 2 choices with our current component impl.: Choice 1: * Add 2 new implementations with hint1 and hint2 (hint1 being what was "default" before) * Keep the implementation with "default" hint but deprecate it and move it to legacy. Refactor it so that it delegates yo hint1 * Add a Provider to decide which impl to use based on whatever conditions we want Choice 2: * Move "default" impl to "hint1" * Add "hint2" * Change "default" implementation to be a Composite that delegates to hint1 or hint2 > But what about those that want to upgrade, but are happy with the older > FOP implementation and don't want to add support for the office connector? Yes, in this case this is transparent for them (and it's a good thing in most caes!). This means they get autoupgrade to something better.

> They'll have to use patched versions of the two component implementations > where their hints are reverted back to the old values. Easy? No. And even > though "default" means "I don't care what the implementation does" It doesn't really mean this. We use "default" only when there's a single implementation. When there are more than 1 each one has a hint that is a qualifier to the implementation since there's no reason one is more default than the other. > , here it does matter a lot which implementation you're using. They have > different requirements, and it's important to know if the implementation > will require an office instance or not. > > Saying that this won't happen since we're all really careful when > designing our components is wishful thinking. We've refactored other more > critical pieces of code than providing alternative implementations for a > component, so we will find ourselves in situations where we'll have to > switch from only one "default" implementation to several. I agree that we've refactored. And it has worked so far right? ;) > Designing our component manager to make it easy to transition is the > right thing to do. > > Still, my major problem is not about overrides, but about the semantics > of "default" (or lack of it). This says nothing about the actual mechanisms > behind the implementation, it just reflects the state of that particular > component implementation: it's the default at the moment. Not caring what > the implementation actually does works only when there is indeed just one > possible implementation that is straight-forward. But in most cases, we do > rely on another library that does the work for us, and libraries die, better > alternatives come along, and changing that internal aspect of the > implementation will sometimes be backwards incompatible, or have a different > behavior. Sure, it does the same job, but it does it so differently that > some will prefer to use the other approach. We have to let users decide > which is their "default", and having multiple implementations with the > "default" hint but different priorities is not very intuitive. Why not make > everything default and remove hints completely if we don't really put any > meaning into the hint?

> > And "default" adds another assumption: XWiki Enterprise is the ultimate > target. Our defaults are the only ones that matter. As an example, all the > *Configuration components have just one "default" implementation, which > relies on xwiki.properties, XWiki.XWikiPreferences etc. Doesn't that tie the > platform to the XWiki Enterprise wiki? This not true in xwiki commons and rendering because I've made sure that we could use them outside of the XWiki Platform. They have default implementation that don't use XWiki Configuration module. > It's not a direct dependency visible at compilation time, it's worse, and > invisible assumption about the final runtime. It's certainly not the default > for other types of users that want to embed xwiki-commons or xwiki-platform > components in a different type of end product. To me this isn't the default > configuration, this is the default configuration used by XWiki Enterprise, > thus my proposal of using something else as the generic component hint > instead of "default". Ok thanks for the explanations. I understand better now what you mean. Actually what you suggest could already be implemented using a best practice of always using Providers when you want a Component injected. In this manner by default you'll get the Generic Provider but anyone could implement a specific Provider implementation for it that would choose between various implementation based on whatever (a value in a META-INF/role-bindings.txt file, data from DB, etc). <side note>Only issue with having Providers everywhere is that you get late verification of your system coherence since dependencies will be resolved only when they're used. OTOH this is a necessity in a fully-dynamic system ;)</side note> Also, the notion of default doesn't always have a meaning. There are lots of cases when there are NO default. For example Macros or Transformations or… Let's continue the discussion it's interesting :) I'd like to review a bit the other Component system out there again to see what they do for this. It's important that they have support for this since we want to be able to switch to them one day. The 3 that I would review are: * Guice * CDI * OSGi

Thanks -Vincent _______________________________________________ devs mailing list devs(a)xwiki.org http://lists.xwiki.org/mailman/listinfo/devs