[IPython-dev] [ATTN: Kernel Authors] libzmq change breaks kernel heartbeat

[MinRK]

ATTN: Kernel authors,

libzmq just released 4.0.6 and 4.1.1. Both of these include a change in
zmq_proxy that breaks the kernel-side of our heartbeat machinery.
Specifically, zmq_proxy cannot be called with a single REP socket for both
ends.

Currently, the message spec docs recommend the kernel-side of the heartbeat
mechanism be implemented as zmq_device(FORWARDER, REP, REP). zmq_device has
wrapped zmq_proxy since libzmq 3, in which case the first argument is
actually ignored.

This can be fixed by making the REP socket a ROUTER, which is what it
should be anyway. There’s one caveat for this change, though: On very old
systems with libzmq-2, ROUTERs don’t work in a FORWARDER device, only
QUEUE. This is fine, and doesn’t really make a difference.

To be fully compatible with libzmq-2.2 through 4.1, the heartbeat call
should change from:

zmq_device(FORWARDER, REP, REP)

to:

zmq_device(QUEUE, ROUTER, ROUTER)

If you only need to support libzmq ≥ 3 (recommended), you can use:

zmq_proxy(ROUTER, ROUTER)

If you were using your own echo loop for the heartbeat socket instead of
zmq_device, you shouldn’t be affected.

This is really an implementation detail, not a protocol change, because it
doesn’t affect how messages are sent or received by the kernel, only the
internal event handling (REP is a restricted subclass of ROUTER,
indistinguishable to peers).

-MinRK
​
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/ipython-dev/attachments/20150605/dd2383ae/attachment.html>