High watermark for durable pub/sub

In the ZeroMQ pub/sub pattern, if we set an identity for a subscriber, the publisher keeps the data in queue till it is actually delivered to the subscriber. That means we could safely disconnect and then reconnect that subscriber without losing data, but it means also that the publisher could easily eat a huge share of memory in the attempt of keeping the required data waiting in its internal queue.

The high watermark is a technique to have some advantage of a durable connection without be forced to pay a too high price in term of memory usage.

It is quite easy to modify a subscriber for durability. Basically we could take the synchronized subscriber we have already written, and specify its identity before connecting it to the publisher:

zmq::context_t context(1);

zmq::socket_t subscriber(context, ZMQ_SUB);

subscriber.setsockopt(ZMQ_IDENTITY, "Hello", 5);

subscriber.setsockopt(ZMQ_SUBSCRIBE, NULL, 0);

subscriber.connect("tcp://localhost:5565");

// ...

If we don't change the synchronized publisher code, we are exposed to the risk of an unlimited growth in memory request. To avoid that, we could determine how many messages we want to store. Say that we want to set it to 2:

zmq::socket_t publisher(context, ZMQ_PUB);

__int64 hwm = 2; // 1

publisher.setsockopt(ZMQ_HWM, &hwm, sizeof(hwm)); // 2

publisher.bind("tcp://*:5565");

1. I am developing for VC++ 2010, as you can see from the 64 bit int type used here.
2. The high watermark has to be set before binding the socket to its endpoint.

Post based on the Z-Guide C durable publisher-subscriber example.