Qpid Proton C++ API  0.12.0

This is the C++ API for the Proton AMQP protocol engine. It allows you to write client and server applications that send and receive AMQP messages.

The best way to start is with the tutorial.

An overview of the AMQP model

Messages are transferred between connected peers over links. The sending end of a link is a proton::sender, and the receiving end is a proton::receiver. Links have named 'source' and 'target' addresses. See "Sources and Targets" below for more information.

Links are grouped in a proton::session. Messages for links in the same session are sent sequentially. Messages on different sessions can be interleaved, so a large message being sent on one session does not block messages being sent on other sessions.

Sessions are created over a proton::connection, which represents the network connection. You can create links directly on a connection using its default session if you don't need multiple sessions.

proton::message represents the message: the body (content), headers, annotations, and so on. A proton::delivery represents the act of transferring a message over a link. The receiver acknowledges the delivery by accepting or rejecting it. The delivery is settled when both ends are done with it. Different settlement methods give different levels of reliability: at-most-once, at-least-once, and exactly-once. See "Delivery Guarantees" below for details.

Sources and targets

Every link has two addresses, source and target. The most common pattern for using these addresses is as follows:

When a client creates a receiver link, it sets the source address. This means "I want to receive messages from this source." This is often referred to as "subscribing" to the source. When a client creates a sender link, it sets the target address. This means "I want to send to this target."

In the case of a broker, the source or target usually refers to a queue or topic. In general they can refer to any AMQP-capable node.

In the request-response pattern, a request message carries a reply-to address for the response message. This can be any AMQP address, but it is often useful to create a temporary address for just the response message.

The most common approach is for the client to create a receiver for the response with the dynamic flag set. This asks the server to generate a unique source address automatically and discard it when the link closes. The client uses this "dynamic" source address as the reply-to when it sends the request, and the response is delivered to the client's dynamic receiver.

In the case of a broker, a dynamic address usually corresponds to a temporary queue, but any AMQP request-response server can use this technique. The server_direct.cpp example illustrates how to implement a queueless request-response server.

Anatomy of a Proton application

To send AMQP commands, call methods on classes like proton::connection, proton::sender, proton::receiver, or proton::delivery. To handle incoming commands, implement the proton::handler interface. The handler receives calls like on_message (a message is received), on_link_open (a link is opened), and on_sendable (a link is ready to send messages).

Messages are represented by proton::message. AMQP defines a type encoding that you can use for interoperability, but you can also use any encoding you wish and pass binary data as the proton::message::body. proton::value and proton::scalar provide conversion between AMQP and C++ data types.

There are two alternative ways to manage handlers and AMQP objects, the proton::container and the proton::connection_engine. You can code your application so that it can be run with either. Say you find the proton::container easier to get started but later need more flexibility. You can switch to the proton::connection_engine with little or no change to your handlers.


proton::container is a reactor framework that manages multiple connections and dispatches events to handlers. You implement proton::handler with your logic to react to events, and register it with the container. The container lets you open multiple connections and links, receive incoming connections and links, and send, receive, and acknowledge messages.

The reactor handles IO for multiple connections using sockets and poll. It dispatches events to your handlers in a single thread, where you call proton::container::run. The container is not thread-safe: once it is running you can only operate on Proton objects from your handler methods.


proton::connection_engine dispatches events for a single connection. The subclass proton::io::socket_engine does socket-based IO. An application with a single connection is just like using proton::container except you attach your handler to a proton::io::socket_engine instead. You can compare examples, such as helloworld.cpp and engine/helloworld.cpp.

Now consider multiple connections. proton::container is easy to use but restricted to a single thread. proton::connection_engine is not thread-safe either, but each engine is independent. You can process different connections in different threads, or use a thread pool to process a dynamic set of connections.

The engine does not provide built-in polling and listening like the proton::container, but you can drive engines using any polling or threading strategy and any IO library (for example, epoll, kqueue, solaris completion ports, IOCP proactor, boost::asio, libevent, etc.) This can be important for optimizing performance or supporting diverse platforms. The example engine/broker.cpp shows a broker using sockets and poll, but you can see how the code could be adapted.

proton::connection_engine also does not dictate the IO mechanism, but it is an abstract class. proton::socket_engine provides ready-made socket-based IO, but you can write your own subclass with any IO code. Just override the io_read, io_write and io_close methods. For example, the proton test suite implements an in-memory connection using std::deque for test purposes.

Delivery guarantees

For at-most-once, the sender settles the message as soon as it sends it. If the connection is lost before the message is received by the receiver, the message will not be delivered.

For at-least-once, the receiver accepts and settles the message on receipt. If the connection is lost before the sender is informed of the settlement, then the delivery is considered in-doubt and should be retried. This will ensure it eventually gets delivered (provided of course the connection and link can be reestablished). It may mean that it is delivered multiple times though.

Finally, for exactly-once, the receiver accepts the message but doesn't settle it. The sender settles once it is aware that the receiver accepted it. In this way the receiver retains knowledge of an accepted message until it is sure the sender knows it has been accepted. If the connection is lost before settlement, the receiver informs the sender of all the unsettled deliveries it knows about, and from this the sender can deduce which need to be redelivered. The sender likewise informs the receiver which deliveries it knows about, from which the receiver can deduce which have already been settled.