[tptacek]

Right, I get that. So I guess my question is: in what ways justifying new terminology are the approaches taken by Urbit different? And whether or not they're different enough from prior art to merit obscure new terms, are those differences worthwhile, or are they different for the sake of being different? Because not all of the last 30 odd years of systems and programming research was pointless.

[pcmonk]

> not all of the last 30 odd years of systems and programming research was pointless.

Indeed, which is why it's a shame our primary server OSes (Unix-likes) can't incorporate them. Urbit can.

As for what are the approaches different, there have been many words spilled over that, but here's a few examples to wit, most of which exist in one or two other systems, but aren't widely used:

- All events are transactions, down to the VM level. There's no concept of an event that left garbage around because power was cut or the machine was rebooted. You can always crash an event, making it as if it never happened.

- Single-level store. Never worry about ORM because your in-memory state never goes away (because all events are transactions).

- Persistent connections with exactly-once messaging. Disconnection is just seen as long latency.

- Strict, purely functional language with a strong type system but no Hindley-Milner (so you don't need category theory).

- Sane indentation for a functional language, known as "backstep".

- The file system is a typed revision control system, which allows intelligent diffs on types other than plain text.

Most, if not all, of these, have been described in research, but nobody's bothered to build a system with them.

[mrkgnao]

Rust uses Hindley-Milner type inference. OCaml uses Hindley-Milner. Swift apparently uses a variant of the system with more features.

None of these make use of monads or anything rooted in category theory.

https://en.wikipedia.org/wiki/Hindley%E2%80%93Milner_type_sy...

HM is a type system that also gives you an algorithm to infer types without having to do the "List<Integer> foo = new ArrayList<Integer>()" dance from languages without type inference. That's it. It has nothing to do with category theory.

[steveklabnik]

(Rust doesn't use Hindley-Milner, though the plan is to try to move closer to it in the future)

[mrkgnao]

Oh, thanks for the correction.

It would appear that my knowledge is out of date: did Rust use something closer to H-M in the past? (Going off of https://github.com/rust-lang/rust/pull/15955 here.)

[steveklabnik]

For more, see here: https://www.reddit.com/r/rust/comments/6fs5q9/language_serve... see aturon's reply about that PR/blog post.

Rust has done almost everything in the past; I wouldn't be surprised if somewhere back in there it was actually HM. :)

[jolux]

Arguably optionals are monadic.

[mrkgnao]

Optionals are also "applicative", but Haskellers used them for a long time before applicative functors became popular. A concept can support a certain way of thinking about it without everyone needing to think that way :)

[jolux]

you don't need to think that way, no, but i was pretty sure that rust optionals were modeled after the maybe monad. perhaps not.

[jolux]

oh ok i misread your original, you were only referring to type inference. i thought you were saying none of the languages mentioned had monads.

[occultist_throw]

> Persistent connections with exactly-once messaging. Disconnection is just seen as long latency.

Bullshit.

https://groups.csail.mit.edu/tds/papers/Lynch/jacm85.pdf

Unless you and Yarvin really think there's a valid answer to the 2 Generals problem... Im sure there's a Fields medal in there if you do.

"Every major message queue in existence which provides any guarantees will market itself as at-least-once delivery. If it claims exactly-once, it’s because they are lying to your face in hopes that you will buy it or they themselves do not understand distributed systems. Either way, it’s not a good indicator."

~ http://bravenewgeek.com/you-cannot-have-exactly-once-deliver...

[pcmonk]

From the second link:

> FLP and the Two Generals Problem are not design complexities, they are impossibility results.

Two generals, in its impossible form, isn't applicable here. For example, this proof[0] depends on the deterministic form being impossible to solve in a finite number of messages. But why would we care about doing it in a finite number of messages? By that measure, at-least-once delivery is impossible too, yet neither you nor the blogger seem to have a problem with that. The solution is easy and common: send a message, ack new messages, retry if you don't receive an ack, never ack an ack, and always ack a dupe. As long as you aren't disconnected forever, this will give you at least once delivery.

FLP only applies in the case of faulty processes. Specifically, FLP arises because the faulty process must either (1) ack the message before handling it, or (2) handle it before acking it. (1) fails if it crashes after acking before handling, causing it to forget it ever received the message (at-most-once). (2) fails if it handles the message, but doesn't persist that fact to disk, so that when the dupe is received it doesn't know it already handled it (at-least-once).

Urbit processes events transactionally (like a database), so you're guaranteed that if you received an ack it's because the other person has persisted that message or its results, and it won't forget about (i.e. they've persisted any state changes related to it). No faulty process, no problem. FLP impossibily doesn't apply.

I suggest actually learning about proposed solutions to problems rather than recalling that someone convinced you they were impossible to solve. Proofs are powerful because they're incontrovertibly true, but they only apply if their conditions are met.

[0] https://en.wikipedia.org/wiki/Two_Generals%27_Problem#Proof

[rhencke]

Urbit claims, like you, exactly-once messaging is impossible: https://news.ycombinator.com/item?id=14666015

But that doesn't mean we can't provide systems where, for all practical purposes, it is a guarantee.

Another way to think about it: SHA1 cannot provide a unique hash for every possible set of content - there are overlaps. This is the pigeonhole principle.

Yet, we have entire systems designed around assuming SHA1 is and will always be a unique hash for a piece of content (Git), and, for all practical purposes, it is a guarantee good enough to work despite that.

[Munksgaard]

I'm curious: What's the advantage of ditching Hindley-Milner?

[Chattered]

I want to give urbit a try. I have a planet, but last time I fired it up the latency was so bad that I wasn't having fun. I'll try again sometime.

I can even get behind runes, and I set up abbreviations in Emacs to learn their names. And generally, I like the philosophy of language where short, one syllable, old/middle English words are used in place of the less familiar Latin/Greek derived words. I don't have a problem with revising our existing vocabulary in this way.

But the attitude to type theory strikes me as anti intellectual, precisely because we get nonsense claims about Hindley Milner requiring category theory. This sort of misinformation does no good for anyone.

[vmchale]

None whatsoever. Hindley-Milner is in fact used widely in language design because of its relatively good error messages (as well as ease of implementation). Also it really doesn't have much to do with category theory.

[vmchale]

>Sane indentation for a functional language, known as "backstep"

This is A) not an impressive problem to solve, and B) not actually a problem in Haskell or Elm.