MIT licensed by Phil Hargett
Maintained by [email protected]
This version can be pinned in stack with:courier-,4742

Inspired by Erlang's simple message-passing facilities, courier provides roughly similar capabilities. Applications simply create one or more endpoints, bind each to a transport using a given name, then can freely send / receive messages to other endpoints just by referencing the name each endpoint bound to its transport.

A primary driver of this package's design is separation of concerns: many algorithms (especially for distributed applications) depend upon a message-passing foundation, but the algorithms are sufficiently complex that the details of how those messages are transmitted among nodes are best hidden away and solved separately from the distributed algorithm itself. With this in mind, this package aims to provide a variety of transports as well as support for common communication idioms (e.g., in order message delivery, selective out of order message delivery, synchronous RPC, etc.).

Applications may process messages in the order received at an endpoint, or use selective message reception to process the first message arriving at an endpoint that also matches a provided selection function. Through selective message reception, applications may approximate the style of an Erlang application, and enjoy better composability of message reception with multiple independent dispatch routines or message pumps.


  • Corrected an issue with memory transport where a message to an unbound destination would be dropped

  • Changed with… functions to return specified types in IO, instead of always returning (); not technically a breaking change, but does prevent backward compatibility from here
  • Memory transport flushes to destinations before completing shutdown
  • Linting and first compile against GHC 8.0.1; continuing to support back to 7.6 for now

  • Updated README with better working example
  • Removed warning about TCP/IP transport; it’s working again

There are significant changes in this release, due to a complete rewrite of TCP connection handling, and, in so doing, a rewrite of the basic model for how transports and endpoints interact. This rewrite fixes bug #10.

Endpoints now simply have inbound and outbound mailboxes, and a set of names to which the endpoint is bound via a transport. The intent is that endpoints are discrete entities orthogonal to the transport which moves messages between them. Its the job of the transport to move messages from the outbound mailbox of one endpoint to the inbound mailbox of the endpoint bound to the name specified in the envelope for a given message.

New helpers for managing the lifecycle of bindings and connections now exist, and also a helper for setting the name of a client.

* Initial implementation of IPv6 transports (TCP and UDP)
* Small cleanup for better thread safety
* Refactored socket transports for more code sharing
* Added anyCall
* Exported RequestId and mkRequestId
* Incorporated changes from carletes for running tests in parallel by allowing
each test to bind to their own set of ports; on machines with multiple CPUs the speedup
has a noticeable impact on producitivity
* Added methodSelector for applications that may want to customize RPC handling

* Included test source into tarball

* Explicitly removed support for 7.4; 7.6 now required.

* Fixed a long-standing issue concerning inconsistent failures. Turns out, on operating
systems with both IPv4 and IPv6 support, the TCP and UDP transports (which were really
only intended to be IPv4 for now) were not properly finding IPv4 addresses--in some
cases, IPv6 were being used instead, resulting in errors.  Both TCP and UDP now
pass (IPv4 only, of course), and in the future, IPv6 versions of these transports
will be added.

* API breaking change: RPC functions no longer take / return instances of Serialize,
but they take / return Message values instead.  The reason for this change is that
while courier does use Serialize internally to simplify a variety of operations,
putting instances of Serialize in RPC functions was misleading: applications would
still need to take care to ensure that different RPC requests / responses could
be differentiated, otherwise courier might accidentally deserialize a message
to the wrong type at an application's request (e.g., inside a hear callback). This
can lead to decoding errors, which may actually causes exceptions to be thrown
in pure methods.  By leaving the interface to use Message values, its not only
consistent with the rest of Endpoint behavior, but it clarifies the responsibility
of the calling application to manage their Message payloads on their own.

* Improvements to RPC. Previously, it was not possible to differentiate correctly
between requests and responses in the same mailbox: the only check was that
deserialization succeeded, which wasn't a sufficient test, and it would be possible
to deserialize a response as a request, etc.  This lead to spurious errors.

* Added functions for detecting presence of messages in an an endpoint's mailbox
(while still leaving them unconsumed): mostly added this for applications that
need to react to the presence of a particular message, but not actual consume
such messages.

* Exposed the Request / Response types used by RPC.  Applications may benefit
from detecting them or otherwise manipulating them outside of normal RPC
flow, they are exposed here for that purpose.

* Added hear primitive as one-shot counterpart to call.  While handle
may be generally useful for many applications as it fully encapsulates repeated
handling of incoming requests, hear adds the ability to wait for a single
incoming request. Hear also returns a function of type Reply suitable for sending
a single response to the caller. Being more low-level in nature, some applications
find direct use of hear more flexible than the higher-level handle construct.

* Fixed a mistake in where RPCs where supposed to have unique identifiers
for matching up a call with a response.  The value was never computed, when it
should have been a random value for every call, so that responses could be matched
with their originating request.  Now using Data.UUID.V4 to generate unique identifiers
for each request; dependency added in cabal.

* Unit tests on OS X 10.9 unreliable again, with both ghc 7.6.3 and ghc 7.8.1 (rc1)

* Unit tests passing on Ubuntu 12.04 LTS AMD64 ghc 7.6.3

* Unit tests passing on FP Complete with 7.8 preview compiler

* Added HelloWorld example, and updated code sample in documentation to reflect
  working with the current API

* Early implementation of synchronous RPCs on top of endpoints, with basic unit tests.
  Both single and group RPCs are implemented, as are calls with definite and indefinite
  waits for responses.

* Removed dependency on network-simple

* Improved documentation throughout.

  • Enabled selective out of order mesage reception, by using a Mailbox, an extension of STM’s TQueues that extracts the next message from the queue that matches a supplied selection function regardless of message order in the queue.

    With selective out of order message reception, endpoints can approximate the Erlang-style of message delivery and composing applications that must handle multiple message types becomes simpler. If an application needs to handle multiple message types, it can choose to run separate separate message pumps concurrently, each handling different types of messages. Each message pump can operate independently of one another.

  • Fixed issue #2 and other interim bugs all resulting from multiple name bindings that resolve the same underlying address. Now, bound sockets are reused (with reference counting) so that if there are multiple bindings to the same address only 1 socket is created and user.

  • Unit tests passing on Mac OS X again, mostly due to correct management of sockets and appropriate reuse.

  • First inclusion of changelog in package

  • Added simple implementation of transport for UDP
  • Verified build & unit tests on Linux Ubuntu 13.10 with GHC 7.6.3
  • Verified build & unit tests on pre-release GHC 7.8 on FP Complete’s platform
  • Mac OS X 10.9 is unreliable for unit tests, do to non-deterministic socket bind failures