conduit 0.5: ResumableSource and finalization

July 22, 2012

GravatarMichael Snoyman

One of the more advanced features of conduit is connect-and-resume. I've discussed the motivation for this elsewhere, but it essentially comes down to simplifying interleaving of input streams. One very practical ramification of this is that the http-conduit API can be drastically simpler than the previous http-enumerator API.

For the most part, connect-and-resume is a technique which hides under the surface, and you don't have to really think about it too much. Starting with conduit 0.5 (and http-conduit 1.5), its semantics have changed a bit and it is now a bit more user-facing, which has led to a bit of confusion. I want to explain why it changed, and how to adapt your code.

Core issue: finalization

At the core of this change sits finalization. Let's take a relatively simple example: reading data from a file. And let's simplify further and ignore exceptions. In the normal conduit workflow (i.e., not using connect-and-resume), the basic workflow is:

  1. Wait till the downstream Sink asks for input.
  2. Open the file handle.
  3. Pull a chunk of data and provide it downstream.
  4. Repeat (3) as long as downstream needs input.
  5. When either downstream is done, or no data is left in the file, close the handle.

Let's ask a simple question: what happens if the Sink never asks for input? We could in theory open the file and immediately close it, but that's inefficient. Instead, we would like to do nothing regarding the file reading in this case. Keep that in mind.

Now, let's take into account connect-and-resume. The basic idea is that, when downstream is done, we don't immediately close the file. Instead, we return a new something (to be defined later) which lets you continue reading the file from where you left off.

something == Source

So we used some kind of "something" above to be the frozen state of the file reading. Let's say for a moment that we wanted this something to be a plain old Source. That works out really nicely: we can reuse all of our existing tools for dealing with Sources for this frozen things.

The problem comes in when this frozen Source is never actually used again. Remember in our first normal connection example, we said that- if the file was never actually read from- no finalization needs to take place. But in our case, the file has already been read from, and therefore we do need to perform finalization, even if the Source isn't actually used.

This was precisely the situation in conduit 0.4. The solution was to always call a Source's finalizer. This means that in some cases, finalizers needed to be defined where they logically did not apply (e.g., how to finalize a sourceFile when you've never read from it). In more technical terms, it means that the PipeM constructor needed to have a finalizer defined as well, which led to lots of redundant code.

conduit 0.5: something == ResumableSource

Starting in conduit 0.5, we've refactored this situation a bit. Finalizers are now associated only when yielding a value downstream. This means that if a Source is never pulled from, no finalizer is registered. This automatically addresses our issue of the unused sourceFile.

The problem comes in with connect-and-resume. In this case, we need to perform some finalization in all cases. To accomodate this, we have a new datatype:

data ResumableSource m o = ResumableSource (Source m o) (m ())

A ResumableSource is just a Source and its associated finalizer. There are three basic operators for interacting with a ResumableSource:

  • $$+ connects a normal Source to a Sink, and returns a ResumableSource.
  • $$++ connects a ResumableSource to a Sink, and returns a new ResumableSource.
  • $$+- connects a ResumableSource to a Sink, and then closes it.

The changes in http-conduit

Hopefully this puts us in a good position to understand what's going on with http-conduit. Simplifying things just a bit, the function to make an HTTP request in http-conduit 1.4 was:

http :: Request -> IO (Source IO ByteString)

In http-conduit 1.5, this changed to:

http :: Request -> IO (ResumableSource IO ByteString)

The reason is that we have a finalizer associated with our Source: we need to shut down our socket, even if we never pull any data from it. In conduit 0.4 and http-conduit 1.4, we could rely on the fact finalizers are always run, even if the Source isn't used. But in 1.5, in order to ensure that the socket is closed, we need to use ResumableSource.

This actually works out to a fairly minor change in usage. In many cases, you can simply replace $$ with $$+-, and continue using the library exactly as you did before.

Getting a Source back

In some circumstances, you may really want to get a Source out of your ResumableSource. (For example, you may have a pre-existing API like wai that you're trying to conform to.) In order to allow this, conduit provides the unwrapResumable function, which gives you both a Source and a finalizer.

But the finalizer and Source it gives you are both specially set up to avoid double-finalization. If the Source is used, then the finalizer doesn't perform any action. On the flip side, if the Source is never used, then the finalizer will perform the originally registered finalization. This way, for instance, your socket will be closed once, and precisely once.

Conclusion

I hope this post clarifies the API changes in http-conduit 1.5. If not, please let me know, and I can try to elaborate a bit more.

Comments

comments powered by Disqus

Archives