[zeromq-dev] defining behavior for running out of FDs

[MinRK]

One of the more common issues I get at pyzmq is how to deal with running
out of FDs.  It is currently not possible to prevent libzmq aborting the
process when the FD limit is exceeded, and inbound connections from peers
are able to bring down whole server processes.

The reason for the abort is, if I recall, that for some of the places where
allocating an FD fails, that failure would result in undefined behavior.

My question is: Can we define the behavior for these failure cases?  If so,
that would eliminate the single most common remaining cause of libzmq
aborts.

I'm not sure where all of the FD requests are, but one is inbound
connections.  Is it possible/sensible to simply reject new connections if
FD allocation fails?

If anyone more familiar with the core can enumerate all of the FD
allocations, then perhaps we can see if it's possible to have defined
failures, rather than unrecoverable crashes.

For pyzmq, it would even be preferable to render a Context and all of its
sockets dead/unusable/closed, rather than aborting the process, but I doubt
that would fly for libzmq devs.

-MinRK
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20120927/eb2ee1ee/attachment.htm>