discovery

[Scott Lystig Fritchie]

>>>>> "uw" == Ulf Wiger <etxuwig@REDACTED> writes:

uw> On 9 Jul 2000, Hal Snyder wrote:

>> We're considering a multicast group address for this purpose. But
>> maybe there's already a solution and I just haven't found it in the
>> docs.
>> Ideally isn't a cluster a fluid thing, without hard-coded physical
>> node assignments?

uw> Ideally, yes. At AXD 301, we have a more flexible (and more
uw> complicated) application controller, called RCM (Runtime
uw> Configuration Manager). Part of it will probably make its way into
uw> OTP, and it will make it easier to configure dynamic clusters on
uw> the fly.

That would be a handy thing.

At the moment, we've been using multicast IP has a hack around this
void.  If a node is rebooted or loses contact with the rest of the
cluster, an alarm eventually goes off and the lost node sends a single
"Anybody out there?  I am 'node@REDACTED'" multicast UDP packet.

The rest of the cluster is also listening for those multicast packets.
(The multicast address is configured out-of-band.)  When one arrives,
a regular Erlang message is sent to 'node@REDACTED', which automagically
triggers the mechanisms used to keep track of Erlang nodes.  I'll
volunteer Jim Larson to speak up if I've misrepresented his work or if
y'all have further questions.  :-)

-Scott