“Unfork” is the opposite of “fork”; whereas forking allows things to run concurrently, unforking prevents things from running concurrently. Use one of the functions in the
Unfork module when you have an action that will be used by concurrent threads but needs to run serially.
A typical use case is a multi-threaded program that writes log messages. If threads use
putStrLn directly, the strings may be interleaved in the combined output.
concurrently_ (putStrLn "one") (putStrLn "two")
Instead, create an unforked version of
unforkAsyncIO_ putStrLn $ \log ->
concurrently_ (log "one") (log "two")
The four async functions are
unforkAsyncIO :: (a -> IO b) -> ( ( a -> IO (Future b) ) -> IO c ) -> IO c
unforkAsyncIO_ :: (a -> IO b) -> ( ( a -> IO () ) -> IO c ) -> IO c
unforkAsyncSTM :: (a -> IO b) -> ( ( a -> STM (STM (Maybe b)) ) -> IO c ) -> IO c
unforkAsyncSTM_ :: (a -> IO b) -> ( ( a -> STM () ) -> IO c ) -> IO c
-- | | | | | |
-- |---------| | |--------------------------| |
-- Original | Unforked action |
-- action | |
These functions all internally use a queue. The unforked action does not perform the underlying action at all, but instead merely writes to the queue. A separate thread reads from the queue and performs the actions, thus ensuring that the actions are all performed in one linear sequence.
There are, therefore, three threads of concern to this library:
- the one running the user-provided continuation
- the one performing the enqueued actions
- the parent thread that owns the other two
Non-exceptional termination works as follows:
- Thread 1 reaches its normal end and halts
- Thread 2 finishes processing any remaining queued jobs, then halts
- Thread 3 halts
Threads 1 and 2 are “linked”, in the parlance of the
async package; if either thread throws an exception, then the other action is cancelled, and the exception is re-thrown by thread 3. Likewise, any exception that is thrown to the parent thread will result in the cancellation of it children. In other words, if anything fails, then the entire system fails immediately. This is desirable for two reasons:
- It avoids the risk of leaving any dangling threads
- No exceptions are “swallowed”; if something fails, you will see the exception.
If this is undesirable, you can change the behavior by catching and handling exceptions. If you want a system that is resilient to failures of the action, then unfork an action that catches exceptions. If you want a system that finishes processing the queue even after the continuation fails, then use a continuation that catches and handles exceptions.
The functions in this module come in pairs: one that provides some means of obtaining the result, and one (ending in an underscore) that discards the action's result.
In the asynchronous case, the result-discarding functions provide no means of even determining whether the action has completed yet; we describe these as "fire-and-forget" functions, because there is no further interaction the initiator of an action can have with it after the action has begun.
The async functions that do provide results are
unforkAsyncIO. Internally, each result is stored in a
MVar, respectively. These variables are exposed to the user in a read-only way:
unforkAsyncSTM gives access to its
STM (Maybe result), whose value is
Nothing while the action is in flight, and
unforkAsyncIO gives access to its
Future result. The
Future type offers two functions:
poll to see the current status (
Nothing while the action is in flight, and
Just thereafter), and
await to block until the action completes.
In both cases, an action is either pending or successful. There is no representation of a “threw an exception” action result. This is because of the “if anything fails, then the entire system fails immediately” property discussed in the previous section. If an action throws an exception, your continuation won't live long enough to witness it anyway because it will be immediately killed.
The two sync functions are
unforkSyncIO :: (a -> IO b) -> IO (a -> IO b )
unforkSyncIO_ :: (a -> IO b) -> IO (a -> IO ())
-- | | | |
-- |---------| |----------|
-- Original action Unforked action
These are much simpler than their asynchronous counterparts; there is no queue, no new threads are spawned, and therefore no continuation-passing is needed. These simply produce a variant of the action that is
bracketed by acquisition and release of an
MVar to assure mutual exclusion.
The hazard of the synchronous approach is that the locking has a greater potential to bottleneck performance.