Chronos - An efficient library for asynchronous programming

Introduction

Chronos is an efficient async/await framework for Nim. Features include:

• Efficient dispatch pipeline for asynchronous execution
• HTTP server with SSL/TLS support out of the box (no OpenSSL needed)
• Cancellation support
• Synchronization primitivies like queues, events and locks
• FIFO processing order of dispatch queue
• Minimal exception effect support (see exception effects)

Installation

You can use Nim's official package manager Nimble to install Chronos:

nimble install https://github.com/status-im/nim-chronos.git

or add a dependency to your .nimble file:

requires "chronos"

Projects using chronos

• libp2p - Peer-to-Peer networking stack implemented in many languages
• presto - REST API framework
• Scorper - Web framework
• 2DeFi - Decentralised file system
• websock - WebSocket library with lots of features

chronos is available in the Nim Playground

Submit a PR to add yours!

Documentation

Concepts

Chronos implements the async/await paradigm in a self-contained library, using macros, with no specific helpers from the compiler.

Our event loop is called a "dispatcher" and a single instance per thread is created, as soon as one is needed.

To trigger a dispatcher's processing step, we need to call poll() - either directly or through a wrapper like runForever() or waitFor(). This step handles any file descriptors, timers and callbacks that are ready to be processed.

Future objects encapsulate the result of an async procedure, upon successful completion, and a list of callbacks to be scheduled after any type of completion - be that success, failure or cancellation.

(These explicit callbacks are rarely used outside Chronos, being replaced by implicit ones generated by async procedure execution and await chaining.)

Async procedures (those using the {.async.} pragma) return Future objects.

Inside an async procedure, you can await the future returned by another async procedure. At this point, control will be handled to the event loop until that future is completed.

Future completion is tested with Future.finished() and is defined as success, failure or cancellation. This means that a future is either pending or completed.

To differentiate between completion states, we have Future.failed() and Future.cancelled().

Dispatcher

You can run the "dispatcher" event loop forever, with runForever() which is defined as:

proc runForever*() =
  while true:
    poll()

You can also run it until a certain future is completed, with waitFor() which will also call Future.read() on it:

proc p(): Future[int] {.async.} =
  await sleepAsync(100.milliseconds)
  return 1

echo waitFor p() # prints "1"

waitFor() is defined like this:

proc waitFor*[T](fut: Future[T]): T =
  while not(fut.finished()):
    poll()
  return fut.read()

Async procedures and methods

The {.async.} pragma will transform a procedure (or a method) returning a specialised Future type into a closure iterator. If there is no return type specified, a Future[void] is returned.

proc p() {.async.} =
  await sleepAsync(100.milliseconds)

echo p().type # prints "Future[system.void]"

Whenever await is encountered inside an async procedure, control is passed back to the dispatcher for as many steps as it's necessary for the awaited future to complete successfully, fail or be cancelled. await calls the equivalent of Future.read() on the completed future and returns the encapsulated value.

proc p1() {.async.} =
  await sleepAsync(1.seconds)

proc p2() {.async.} =
  await sleepAsync(1.seconds)

proc p3() {.async.} =
  let
    fut1 = p1()
    fut2 = p2()
  # Just by executing the async procs, both resulting futures entered the
  # dispatcher's queue and their "clocks" started ticking.
  await fut1
  await fut2
  # Only one second passed while awaiting them both, not two.

waitFor p3()

Don't let await's behaviour of giving back control to the dispatcher surprise you. If an async procedure modifies global state, and you can't predict when it will start executing, the only way to avoid that state changing underneath your feet, in a certain section, is to not use await in it.

Error handling

Exceptions inheriting from CatchableError are caught by hidden try blocks and placed in the Future.error field, changing the future's status to Failed.

When a future is awaited, that exception is re-raised, only to be caught again by a hidden try block in the calling async procedure. That's how these exceptions move up the async chain.

A failed future's callbacks will still be scheduled, but it's not possible to resume execution from the point an exception was raised.

proc p1() {.async.} =
  await sleepAsync(1.seconds)
  raise newException(ValueError, "ValueError inherits from CatchableError")

proc p2() {.async.} =
  await sleepAsync(1.seconds)

proc p3() {.async.} =
  let
    fut1 = p1()
    fut2 = p2()
  await fut1
  echo "unreachable code here"
  await fut2

# `waitFor()` would call `Future.read()` unconditionally, which would raise the
# exception in `Future.error`.
let fut3 = p3()
while not(fut3.finished()):
  poll()

echo "fut3.state = ", fut3.state # "Failed"
if fut3.failed():
  echo "p3() failed: ", fut3.error.name, ": ", fut3.error.msg
  # prints "p3() failed: ValueError: ValueError inherits from CatchableError"

You can put the await in a try block, to deal with that exception sooner:

proc p3() {.async.} =
  let
    fut1 = p1()
    fut2 = p2()
  try:
    await fut1
  except CachableError:
    echo "p1() failed: ", fut1.error.name, ": ", fut1.error.msg
  echo "reachable code here"
  await fut2

Chronos does not allow that future continuations and other callbacks raise CatchableError - as such, calls to poll will never raise exceptions caused originating from tasks on the dispatcher queue. It is however possible that Defect that happen in tasks bubble up through poll as these are not caught by the transformation.

Platform independence

Several functions in chronos are backed by the operating system, such as waiting for network events, creating files and sockets etc. The specific exceptions that are raised by the OS is platform-dependent, thus such functions are declared as raising CatchableError but will in general raise something more specific. In particular, it's possible that some functions that are annotated as raising CatchableError only raise on some platforms - in order to work on all platforms, calling code must assume that they will raise even when they don't seem to do so on one platform.

Exception effects

chronos currently offers minimal support for exception effects and raises annotations. In general, during the async transformation, a generic except CatchableError handler is added around the entire function being transformed, in order to catch any exceptions and transfer them to the Future. Because of this, the effect system thinks no exceptions are "leaking" because in fact, exception handling is deferred to when the future is being read.

Effectively, this means that while code can be compiled with {.push raises: [Defect]}, the intended effect propagation and checking is disabled for async functions.

To enable checking exception effects in async code, enable strict mode with -d:chronosStrictException.

In the strict mode, async functions are checked such that they only raise CatchableError and thus must make sure to explicitly specify exception effects on forward declarations, callbacks and methods using {.raises: [CatchableError].} (or more strict) annotations.

Multiple async backend support

Thanks to its powerful macro support, Nim allows async/await to be implemented in libraries with only minimal support from the language - as such, multiple async libraries exist, including chronos and asyncdispatch, and more may come to be developed in the futures.

Libraries built on top of async/await may wish to support multiple async backends - the best way to do so is to create separate modules for each backend that may be imported side-by-side - see nim-metrics for an example.

An alternative way is to select backend using a global compile flag - this method makes it diffucult to compose applications that use both backends as may happen with transitive dependencies, but may be appropriate in some cases - libraries choosing this path should call the flag asyncBackend, allowing applications to choose the backend with -d:asyncBackend=<backend_name>.

Known async backends include:

• chronos - this library (-d:asyncBackend=chronos)
• asyncdispatch the standard library asyncdispatch module (-d:asyncBackend=asyncdispatch)
• none - -d:asyncBackend=none - disable async support completely

none can be used when a library supports both a synchronous and asynchronous API, to disable the latter.

TODO

• Pipe/Subprocess Transports.
• Multithreading Stream/Datagram servers

Contributing

When submitting pull requests, please add test cases for any new features or fixes and make sure nimble test is still able to execute the entire test suite successfully.

chronos follows the Status Nim Style Guide.

Other resources

• Historical differences with asyncdispatch

License

Licensed and distributed under either of

• MIT license: LICENSE-MIT or http://opensource.org/licenses/MIT

or

• Apache License, Version 2.0, (LICENSE-APACHEv2 or http://www.apache.org/licenses/LICENSE-2.0)

at your option. These files may not be copied, modified, or distributed except according to those terms.