[zeromq-dev] RFC: Locating 0mq services

[Martin Lucina]

Hi all,

those of you who have used 0mq 1.x will know that it came with a
"zmq_locator" daemon that could be used to locate resources using symbolic
names. Over the last week or so I've been discussing some ideas with Martin
Sustrik and Palo Malosek about how to cheaply (in terms of development
time) get this functionality for 0mq 2. Here are the two most interesting
proposals:

1) Using DNS TXT records to locate services:

This could be implemented in a quite straightforward way as follows:

- Implement a new "transport" called 'locator' which would use URIs with
  the syntax 'locator://service.name'.

- When used with zmq_bind() or zmq_connect() the transport would look up
  the TXT record for 'service.name' in the DNS. The TXT record data would
  contain the actual URL to use (e.g. tcp://hostname:port) which would then
  be resolved using the existing code.

- In future, multiple TXT records for a single service name could be used
  to implement e.g. round robin connection to multiple servers or failover.

Implementation notes:

This would require using the appropriate system-dependent resolver API,
which would mean at least separate code for Win32 (using the DNS API at
http://msdn.microsoft.com/en-us/library/ms682100(VS.85).aspx) and *NIX
systems (using resolver(3)).

Note that the 'dns' transport would have to be optional for systems which
do not provide/for which we don't yet implement the DNS resolver.

2) Using a simple static file:

This option would be useful for those users who don't want to or can't
change their DNS infrastructure. Rather than implementing a service similar
to zmq_locator in 1.x we'd use a simple file.

The easiest way to implement this would be to define an environment
variable such as ZMQ_LOCATOR. The user would set this to point to an XML
configuration file which would be loaded at zmq_init() time and could look
something like this:

<locator>
  <service name = "feed1">
    <connect addr = "tcp://10.10.10.1:5555" />
    <!-- for failover -->
    <connect addr = "tcp://10.10.11.1:5555" />
    <!-- for services binding to this name as opposed to connecting -->
    <bind addr = "tcp://eth0:5555" />
  </service>
</locator>

Again, the same 'locator://service.name' syntax would be used. The
'locator' transport would first look up the service in the ZMQ_LOCATOR
file if present and then DNS (also if present/configured).

Both of the options I describe are very low cost to implement and do not
add any additional third party dependencies to 0mq.

Thoughts?

-mato