Big Subsite Rewrite

March 15, 2013

GravatarMichael Snoyman

In my previous blog post, I alluded to some more radical changes I'd been playing with for the Yesod 1.2 release. Initially, I thought I wouldn't be including them for now, but after some experimentation, I think I'm going to go ahead and include this.

For a quick overview of what's happening, you can look at the updates I've made to the following codebases:

I've also been keeping the detailed changelog up-to-date with the coming changes. And if you just want to jump ahead and see what a subsite looks like, go the the end of the blog post.


If it ain't broke, don't fix it. So let's ask the question: what's wrong with Yesod 1.1, and in particular the subsite integration? The following are some annoyances I've had with it for quite a while:

  • GHandler and GWidget both take type parameters for both the subsite and the master site, even if you aren't writing a subsite.
    • This added complication appears everywhere.
    • It's unclear whether users should write code that says GHandler sub App () or GHandler App App () (the latter having the convenient synonym Handler ()).
    • If you use the latter, your code is more readable, but may cause problems if you later decide to use a subsite.
    • Even if you use the prettier Handler wrapper, error messages will still mention GHandler App App.
  • Speaking of error messages, we had really ugly onces back in the old days of the GHandlerT monad transformers (anyone remember GHandlerT sub App (Control.Monad.Trans.Resource.ResourceT IO)?). To work around this, GHandler and GWidget are no longer monad transformers.
    • But we still want to be able to pretend that they're transformers, so we have a custom lift function.
    • We can't do useful things like restrict the actions a handler can do by changing its underlying monad.
  • We have this weird arbitrary encoding of master and subsite. What if a subsite wants to have a subsite? Well, it's possible, but it's quite confusing how that fits in with datatypes that only mention two site parameters.
  • Many functions (like getCurrentRoute) refer to the subsite instead of the master site. As a result, you need to play around with things like getRouteToMaster even if you're not writing a subsite.

And then of course the main issue everyone probably has with subsites: they're too confusing to create! Actually, I think this is mostly a documentation issue, with a little bit of Template Haskell woes thrown in. We'll get to that later, but for now I want to talk about how we're going to change the core of Yesod itself to allow for better subsites.

Back to transformers

Let's dive in with the new and improved approach:

newtype HandlerT site m a
newtype WidgetT  site m a

That's basically the whole change I'm talking about. Instead of a subsite and master site parameter, we have a single site parameter. And instead of hard-coding a specific underlying monad, the monad is a parameter again. This immediately solves all the custom lift nonsense. And clearly, if you're not writing subsite code, you won't even have to think about subsites, because there is no subsite!

So now, if I'm writing the application App, my handler functions look like:

myAppHandler :: HandlerT App IO RepHtml

And just like now, we'll have the convenient Handler type synonym, now defined as:

type Handler = HandlerT App IO

So for the majority of user code just using the Handler synonym, nothing has to change.

But that just begs the question: how do you write code that lives in a subsite? Easy: you stack the transformers. Suppose you're working in the Auth subsite; a handler function would look something like this:

myAuthHandler :: HandlerT Auth (HandlerT App IO) RepHtml

Or more generally, you could allow myAuthHandler to work with any master site:

myAuthHandler :: HandlerT Auth (HandlerT master IO) RepHtml

Or you could require that the master site implement some kind of interface via typeclasses:

myAuthHandler :: YesodAuth master => HandlerT Auth (HandlerT master IO) RepHtml

There's no longer any confusion about "which site" a function lives in, since HandlerT only knows about a single site. So getCurrentRoute just returns a route, not a "subsite route" or a "master site route". getYesod returns the site foundation type, not sub or master site.

But if you're writing a subsite, you might need to access information from the parent. But that's now trivial: just use lift. For example:

myAuthHandler :: HandlerT Auth (HandlerT App IO) RepHtml
myAuthHandler = do
    auth <-      getYesod -- returns an Auth
    app  <- lift getYesod -- returns an App

This change requires some code rewriting for upgrading, but after working on it a while I believe that the trade-off is worth it.

Where's ResourceT?

If you're paying close attention, you might be wondering where ResourceT went. The answer is that it's actually embedded inside of HandlerT. In fact, I just added some code to resourcet to make this kind of usage more efficient. The reasons I went for this embedding are:

  • Avoiding an extra transformer layer can mean increased performance.
  • Type signatures and error messages will stay simpler.

This means that, unlike Yesod 1.1, using lift inside a Handler will not let you run a ResourceT IO action. Instead, you'll want to use liftResourceT.

And what about WidgetT

I've followed almost exactly the same formulation for WidgetT. And if you're not writing subsites, there's really only one other thing you need to know. Instead of lifting handler actions into your WidgetT, most handler actions live in typeclasses now, so they can be automatically used in a Widget. If you do have a Handler function that you want to lift, you'll need to use handlerToWidget. (That might get changed to liftHandlerT instead, but that's a different discussion...)

For subsites, the situation is a little bit more tricky. To understand why, consider the new type signature for the defaultLayout method:

defaultLayout :: WidgetT site IO () -> HandlerT site IO RepHtml

Now consider that you're writing a widget in a subsite that refers to a route in the subsite. This might look something like:

[whamlet|<a href=@{LoginR}>Please login|]

Since LogingR is a route in the Auth subsite, the type of that widget is WidgetT Auth IO (). But if we're working on the App site, defaultLayout expects a WidgetT App IO (). How do we reconcile the two? I've come up with two approaches.

  1. Never created subsite widgets. Instead, whenever you want to embed a subsite URL, you convert it to a master site URL. Then, you can apply defaultLayout and lift its result: For example:

    toParent <- getRouteToParent
    lift $ defaultLayout [whamlet|<a href=@{toParent LoginR}>Please login|]
  2. Provide a helper function to lift up a widget into the parent site. This looks something like:

    liftWidget :: WidgetT child IO a
               -> HandlerT child (HandlerT parent m) (WidgetT parent m a)

    Then our example would be:

    widget <- liftWidget [whamlet|<a href=@{LoginR}>Please login|]
    lift $ defaultLayout widget

So far, I favor the first, as it's pretty close to what we do already. This still doesn't feel as elegant as I'd like it to be. However, in its favor, this approach doesn't really perform any magic: it's fairly obvious what each step is doing. And for the most part, this will be boilerplate stuff that all subsites use, so it should be something that can be learnt once and reused.

Putting it together

So how do you actually create a subsite? It's much like creating a normal site, with a few differences:

  1. Use mkYesodSubData to create your route datatype and rendering function.
  2. Create a YesodSubDispatch instance, using the mkYesodSubDispatch TH function to generate the dispatch function itself.
  3. Instead of a plain Handler, you'll have a Handler stack, something like: HandlerT SubSite (HandlerT master IO).

I've put together a subsite demo which demonstrates how to create a subsite. Here are some important excerpts from that demo.

Creating routes is very similar to how you do so for a master site. Instead of mkYesod, you use mkYesodSubData:

mkYesodSubData "Wiki" [parseRoutes|
/ WikiHomeR GET
/read/*Texts WikiReadR GET
/edit/*Texts WikiEditR GET POST

To set up dispatch, you must create an instance of YesodSubDispatch, using the mkYesodSubDispatch TH helper function to generate the actual code. You can put whatever restrictions you want on the master site by changing the typeclass constraints.

instance YesodWiki master => YesodSubDispatch Wiki (HandlerT master IO) where
    yesodSubDispatch = $(mkYesodSubDispatch resourcesWiki)

As described above, you have to remember to lift some functions and convert subsite routes to the parent site, e.g.:

toParent <- getRouteToParent
lift $ defaultLayout
            <a href=@{toParent $ WikiReadR page}>Read page

There is definitely still some room for improvement here. We can probably find some common patterns to be abstracted out, like "apply defaultLayout and convert all the routes". But overall, this approach feels much cleaner to me than what we have currently. And most importantly, it takes the complexity out of the majority of apps entirely.

In my opinion, the only downside with this change is its breaking nature, but hopefully most breakage can be mechanically fixed (e.g., replace GHandler Foo Foo with HandlerT Foo IO). If people see other problems, or think I'm understating the effect of the breakage, please bring it up.


comments powered by Disqus