Conduit.3.0.0

[dinosaure]

Hi all,

I think it's time to start to discuss about a possible migration of conduit to the code-base of tuyau. More generally, one year ago we decided to find a proper way to replace conduit by something more powerful.

conduit can be hard to understand and it's why I wrote an article to explain more precisely what I had in my mind: Tuyau, the next conduit.

Some tasks come to my mind:

• be able to properly initialize a TLS connection (certificates, etc.)
• a better interface about Service
• be able to extend outside the library protocols available

I think we solved all of that into a (not so) new interface. The key of this deep update is the use of extensible variant and GADTs.

So I think it's time to talk more technically about a possible migration, details about implementation, some missing points about comprehension (which is possible even if I wrote an article about that) and finally a plan over all MirageOS projects to migrate to the new interface.

I really would like to aggregate opinions of all people where conduit is everywhere and I would not like to spoil anybody about a possible migration.

The best of what you should do is to read my article (which wants to give you a global view of the project) and examples available on my repository to really understand what is going on. For an other feedback, my current blog and https://paste.x25519.net/ uses both tuyau instead to use conduit as a way to initialize a server (I still continue to use conduit for irmin however as a client). I did not get crashes at this layer from what I know.

I'm really disposal to talk about some points:

• the idea to use [ host ] Domain_name.t instead Uri.t
• enforce at the first layer type of input and output
• keep the ability to switch between Lwt or Async (or Unix as my example)
• the idea to use a Map.t (no side-effect) instead an Hashtbl.t as the global resolver
• the new interface of FLOW (eg. val read : flow -> input -> ([ Eof | Input of int ], error) result)
• the new interface of SERVICE (eg. listen & accept)
  about that, you mostly be interested by tuyau_mirage_tcp which re-implement accept to
  be close to Unix.accept
• the way to compose a FLOW with TLS (eg. tls/tuyau_tls.ml is the main implementation for any)
• the global design which expects that the user must fill explicitly the Map.t to be able to resolve instead that a filling my application of functor (as conduit does with functoria)

So don't hesitate to participate, go deeply into my code, improve the documentation and share your opinion about that 👍 ! I will let this issue open for 2 weeks, then, depending on what we said, I will do what we will agree. Thanks!

[nickbetteridge]

I'm an absolute fan of Uri.t 👍

[dinosaure]

After many experimentation on both side (client and server), I think it is the time to start to integrate tuyau into conduit now. I will necessary break the API but it's for better things.

[avsm]

Sounds good. I'm happy to make the first iteration Lwt only, followed by a separate Async one later.

[dinosaure]

See #311