A lightweight structured concurrency library.
A lightweight structured concurrency library.
For a variant of this API generalized to MonadUnliftIO
, see ki-unlifted
.
Remember to link your program with -threaded
to use the threaded runtime!
ki | ki-unlifted |
---|---|
Overview
ki
is a lightweight structured-concurrency library inspired by many other projects and blog posts:
- libdill
- trio
- Kotlin coroutines
- Notes on structured concurrency, or: Go statement considered harmful
- Structured Concurrency in High-level Languages
- Update on Structured Concurrency
- Two Approaches to Structured Concurrency
- libdill: Structured Concurrency for C
A previous version of ki
also included a mechanism for soft-cancellation/graceful shutdown, which took inspiration from:
- Go Concurrency Patterns: Context
- .NET 4 Cancellation Framework
- Timeouts and cancellation for humans
- Graceful Shutdown
However, this feature was removed (perhaps temporarily) because the design of the API was unsatisfactory.
Documentation
Example: Happy Eyeballs
The Happy Eyeballs algorithm is a particularly common example used to demonstrate the advantages of structured concurrency, because it is simple to describe, but can be surprisingly difficult to implement.
The problem can be abstractly described as follows: we have a small set of actions to run, each of which can take arbitrarily long, or fail. Each action is a different way of computing the same value, so we only need to wait for one action to return successfully. We don't want to run the actions one at a time (because that is likely to take too long), nor all at once (because that is an improper use of resources). Rather, we will begin executing the first action, then wait 250 milliseconds, then begin executing the second, and so on, until one returns successfully.
There are of course a number of ways to implement this algorithm. We'll do something non-optimal, but simple. Let's get the imports out of the way first.
import Control.Concurrent
import Control.Monad (when)
import Control.Monad.STM (atomically)
import Data.Function ((&))
import Data.Functor (void)
import Data.List qualified as List
import Data.Maybe (isJust)
import Ki qualified
Next, let's define a staggeredSpawner
helper that implements the majority of the core algorithm: given a list of actions, spawn them all at 250 millisecond intervals. After all actions are spawned, we block until all of them have returned.
staggeredSpawner :: [IO ()] -> IO ()
staggeredSpawner actions = do
Ki.scoped \scope -> do
actions
& map (\action -> void (Ki.fork scope action))
& List.intersperse (threadDelay 250_000)
& sequence_
atomically (Ki.awaitAll scope)
And finally, we wrap this helper with happyEyeballs
, which accepts a list of actions, and returns when one action returns successfully, or returns Nothing
if all actions fail. Note that in a real implementation, we may want to consider what to do if an action throws an exception. Here, we trust each action to signal failure by returning Nothing
.
happyEyeballs :: [IO (Maybe a)] -> IO (Maybe a)
happyEyeballs actions = do
resultVar <- newEmptyMVar
let worker action = do
result <- action
when (isJust result) do
_ <- tryPutMVar resultVar result
pure ()
Ki.scoped \scope -> do
_ <-
Ki.fork scope do
staggeredSpawner (map worker actions)
tryPutMVar resultVar Nothing
takeMVar resultVar