Concurrency In Candy

How Candy uses structured concurrency

by Marcel Garus · 2023-12-4
available at www.marcelgarus.dev/candy-concurrency

As of 2023-08-10, concurrency in Candy works differently. See this article for the reasoning behind that.

Many programming languages have the option to spawn a new thread. A downside of this is that the control-flow becomes non-local: If you call a function, you can't see if it starts another thread that then runs in the background.

sin(2// may spawn a thread? who knows?

Candy is a functional language and we want functions to be a perfect unit of abstraction – you should know be able to reason about a function call without looking into the function's source code. That's why we opted to use structured concurrency.

concurrency with threads vs. structured concurrency

The idea of structured concurrency is to enforce that spawned control flows are joined eventually. In Candy, that happens with a new kind of scope, a parallel scope. Inside, you can use the given nursery to spawn new threads of execution, called fibers. Only once all fibers spawned on the nursery completed, does the parallel section itself end.

parallel nursery ->
  async nursery print "Banana" }
  async nursery print "Grapefruit" }
}
print "Kiwi"

Here, parallel starts a new parallel scope. Inside, two fibers are spawned using the async function with the nursery. There's no guarantee about the order of execution, so this code prints "Banana" and "Grapefruit" in any order. Only once both fruits are printed, does the parallel section end and the code print "Kiwi".

There's also an await function, the counterpart to async. That way, you can model dependencies between spawned fibers.

parallel nursery ->
  willBeOne async nursery {
    print "One"
    1
  }
  async nursery print "Two" }
  one await willBeOne

  # async and await cancel each other out
  three await async }
}

Channels

Concurrently executing fibers are only one part of the equation. They can communicate using channels. Channels can store a number of values – you can send values to them on one end and receive them from the other end.

channels
channel.create 4  # capacity of 4
tx c.sendPort
rx c.receivePort

send tx 1
send tx 2
send tx 3

one receive rx

How does it work?

The Candy runtime maintains a list of fibers and a list of channels. It then chooses a random running fiber and continues executing instructions. If a fiber starts a parallel section or interacts with a channel, it's paused.

fibers

In this example, fiber 1 started a parallel section and spawned three children. One of the fibers called receive on a channel. Because there are no values, it's waiting for items to be sent to the channel. The runtime will randomly choose between fiber 2 and 4 and run either of them.

Well, but…

Actually, this concurrency approach is a thing of the past. As of PR X, we removed concurrency from the Candy runtime. It's no longer a builtin concept, but may be implemented in the underlying platform where Candy is embedded. For more information about the reasoning behind that change, see this article.