[Haskell-cafe] Can you do everything without shared-memory concurrency?

[Jason Dusek]

Bruce Eckel <bruceteckel at gmail.com> wrote:
> ...shared-memory concurrency is impossible for programmers to
> get right...

  Explicit locking is impractical to get right. Transactional
  interfaces take much of the pain out of that -- even web
  monkeys can get shared memory right with SQL!

  When two instances of a web app interact with the database,
  they are sharing the database's memory. So you have locking,
  mutual corruption and all that jazz -- yet you seem to be
  message passing! More generally, applications backed by
  network services -- which are presented through a message
  passing interface -- are shared memory applications much of
  the time (though a bright service can tell when modifications
  are unrelated and run them in parallel).

  Message passing certainly makes it easier to write parallel
  applications, but does it provide any help to manage shared
  state? No. None whatsoever. If you have a bunch of programs
  that never share state with one another, they are a single
  application in name only.

  Using locking as your default mode of IPC is harrowing, but
  you can do anything with it. Using message passing is simpler
  in most cases, but you'll have to implement locking yourself
  once in awhile. Language level, transactional interfaces to
  memory are going to cover all your bases, but are rare indeed
  -- as far as I'm aware, only Haskell's STM offers one.

> ...the unnatural domain-cutting that happens in shared-memory
> concurrency always trips you up, especially when the scale
> increases.

  This is true even with transactional interfaces. Message
  passing is _like the network_ and makes you think about the
  network -- so when it's time to get two servers and hook them
  together, you are already ready for it already. Transactional
  shared memory is not at all like the network. Why not? In a
  transactional system, a transaction can not both be approved
  and unwritten. On the network, though, these are separate
  messages, going in different directions -- they can fail
  independently.

--
_jsn