Disemboweling WAI (aka gutting out conduit)

April 28, 2014

GravatarMichael Snoyman

The Haskell Web Application Interface- or WAI- serves as a low-level interface between web applications and servers. In order to do this in a resource-efficient manner, it avoids lazy I/O and instead uses explicit streaming. The story of how it does that streaming has evolved over time: it initially used its own home-brewed streaming abstraction. Later, Gregory Collins convinced me to switch over to enumerator. In the 1.0 release, it switched to conduit, which was essentially designed for the very purpose of supporting WAI's use cases. While I'm very happy with conduit, baking conduit into WAI makes the barrier to using a different streaming framework (like pipes) higher.

So today, I'm proposing that we go all the way back to the beginning, and remove dependencies on external streaming frameworks from WAI, making it a completely universal web interface for Haskell.

I've been hesitant about doing this in the past due to two different reasons:

  1. Making this change makes it more difficult to write handlers and middleware for WAI.
  2. It's not really possible to get rid of a streaming abstraction entirely; instead, we end up just having a locally baked abstraction, which is less thoroughly tested than existing abstractions.

On the first point, most middlewares and handlers which modify request and response bodies are already maintained by the WAI team (and mostly by me personally), so I'm less concerned about pushing such a burden out onto the community. Thankfully, applications and frameworks can be completely insulated by this change by providing a wai-conduit adapter package.

Regarding the second point, I've recently had some experience with this: refactoring http-client out of http-conduit. It turned out to be relatively painless, though I did end up having to reimplement a number of combinators from conduit (especially in the test suite). Nonetheless, the codebase is about the same level of complexity, given the low-level nature of the http-client library, and there's been an overwhelmingly positive response to the splitting up of those two packages, so I want to try it out with WAI as well.

I've created a no-conduit branch in the WAI repo. Currently, I've converted the wai and warp repos over to be conduit-free (with a few helper functions stubbed out). And Warp passes its full test suite, which is rather promising.

The streaming interface is relatively simple. In order to consume a request body, you use the function:

requestBody :: Request -> IO ByteString

This function returns the next strict ByteString from the request body in the stream, or an empty ByteString if the body has been fully consumed. On the response side, you write an application with the type:

(Maybe Builder -> IO ()) -> IO ()

The argument function is used for emitting data to the user. If you provide a Just value, it sends the data in, and a Nothing value flushes the stream. (Currently, WAI uses the Flush data type from conduit for this purpose.)

The code is not yet ready to be released, but it is ready for some review and discussion. I'm hoping to hear community feedback, both from current users of WAI, and those who are considering using it (either directly in an application, or as part of a framework).


A separate breaking change that came up when working on this feature has to do with safe resource allocation and the definition of Application. In WAI 2.0, we introduced the function responseSourceBracket, which is a version of bracket that works for WAI applications. Internally to Warp (and every other WAI handler), we had to jump through some hoops to make sure async exceptions were all masked and restored properly. We also had to make our definition of ResponseSource a bit ugly to make this all possible.

Now with the move away from Sources, we have two choices: either perpetuate the strangeness in the ResponseStream and ResponseRaw constructor, or add bracket semantics to the entirety of a WAI application. In terms of code, I'm talking about replacing:

type Application = Request -> IO Response

with

type Application = Request -> (forall b. (Response -> IO b) -> IO b)

I've implemented this idea as well. It certainly makes the Warp codebase easier to follow, and allows for some slightly more powerful applications (e.g., acquiring a resource for use in some lazy I/O performed by ResponseBuilder). The downsides are:

  • Yet another breaking change.
  • It's harder to explain the intuition versus the dead simple Application we have now.
  • It will likely make middlewares significantly harder to write, though I'll admit that I haven't tried that yet.

When I discussed this change with Gabriel, he pointed out his Managed data type as a possible approach to make the signatures and usage slightly less scary. That would mean:

newtype Managed r = Managed { _bind :: forall x . (r -> IO x) -> IO x }
type Application = Request -> Managed Response

Managed would be nice in that it provides a number of instances (including Monad), but it makes the WAI API a little bit denser to grok. I think I'm leaning against that direction, but wanted to raise this in the discussion as well.

Comments

comments powered by Disqus

Archives