Drop Conduit's Finalizers?

tl;dr: I'm thinking of dropping finalizers from Conduit. If you have use cases that necessitate finalizers, please share them with me. Github issue for examples.

People who have followed the history of the Conduit library will likely be aware of how passionately I've argued for the necessity of a finalizer concept. For those unaware: when you have a Conduit pipeline like the following:

runConduitRes $ do
  sourceFile input .| takeCE 100 .| sinkFile output
  -- lots of other stuff

Without finalizers, the file input will remain open until we exit the runConduitRes call (unless the file is less than 100 bytes). The reason is that, in a pipeline, downstream is always in control. When takeCE stops consuming from sourceFile, sourceFile will never be called again. Finalizers allow sourceFile to tell downstream "hey, if you don't call me back, please close up my file." And so they're a Good Thing.

They're also a Complicated Thing. The internals of Conduit include a lot of complexity around finalizers. As a user, you hardly ever see it, which is by design. But if you start dealing with the guts of Conduit, it's there. (Have a look at this commit to get an idea.)

They also complicate the story around associativity of Conduit significantly. Or more to the point: without finalizers, we can get properly behaving associative and identity laws*. With finalizers, we have to start making up claims about reordering of finalizers. Which is almost always fine in practice, but clearly not a mathematical law.

* Caveats: I've never written the proof of it completely. Also, it relies on using the type paramter on the Pipe type to eliminate leftovers, but leftovers are not a topic I'm raising right now.

None of this is new information; so why am I writing this blog post now? The first is that I'm already working on a breaking change to Conduit to standardize naming and eliminate some legacy type and operator names. See the discussion and initial comparison for more details. This naturally leads me to ask related questions.

More to the point: after having worked with Conduit for years, my initial concerns about prompt finalization seem to have been overzealous. While the code above can happen, it doesn't happen that often in practice, and even that level of resource overholding isn't usually that bad. Regardless, if the situation really does call for guaranteed promptness, we can still get it (a trick I'm fairly certain I learned from Gabriel Gonzalez):

runConduitRes $ do
  withSourceFileResource input $ \src ->
    src .| takeCE 100 .| sinkFile output
  -- lots of other stuff

It's not quite as elegant, but works well enough. And again: I'm hard pressed to think of real life code I've written that would really warrant this. BIG REQUEST If you have examples of Conduit-using code where finalizers are vital, please send me examples.

While on the one hand, making this change would be to admit I've sucked up huge amounts of extra work in maintaining Conduit over the years, I'd be very happy to cut the dead weight now, unless someone out there wants to convince me otherwise. Feel free to discuss wherever desired, but the main discussion I'll be sure to follow will be conduit issue #343.

There are some natural follow-on questions that come from this also, which I'll likely broach in a later blog post. To give a taste and hopefully encourage some thoughts from others:

• Should the Pipe types upstream finalizer concept disappear as well?
• Can we remove the Leftover data constructor from Pipe, making Pipe a full category, and then move the tracking of leftovers to ConduitM and its codensity approach?