Kenneth,
Thank you for the thorough response! Your reasoning is appealing and I'd
like to engage in the discussion on how we can move on to include this
functionality in the distribution, as I believe the community will benefit
greatly from being able to run an Erlang node listening on multiple transports
concurrently.
The changes in the patch (https://github.com/erlang/otp/pull/121) cover the
following areas:
1. Extension of the EPMD distributed protocol (here we can safely add new
set of commands while being fully backward compatible)
erts/doc/src/erl_dist_protocol.xml
2. Modification of the EPMD daemon to support the new protocol version and
to keep track of the multiple protocols/ports per node (this is also done
in a backward compatible way. Nodes supporting new protocol version
communicate using new spec, and old nodes communicate using old spec)
erts/epmd/src/epmd.h
erts/epmd/src/epmd.c
3. Modification of the EPMD command-line tool to talk to the local EPMD
daemon using the new protocol version (the command-line tool uses new
protocol version. This is reasonable, since it's installed together with
the version of VM supporting the new EPMD protocol features)
erts/epmd/src/epmd_cli.c
erts/epmd/src/epmd_cli.c
erts/epmd/src/epmd_int.h
erts/epmd/src/epmd_srv.c
4. Updates of the UDS distribution example code. I believe the example in
the current release is outdated (I was unable to compile it), but I patched
it anyway to illustrate the usage of the new protocol)
lib/kernel/examples/uds_dist/src/dist_selector.erl
lib/kernel/examples/uds_dist/src/uds_dist.erl
lib/kernel/examples/uds_dist/src/uds_server.erl
5. Change of the net_kernel transport pluggable architecture to allow
support of the new EPMD protocol.
5.1 Net_kernel support for the new protocol.
5.2 TCPv4 support of the new protocol
5.3 TCPv6 support of the new protocol
5.4 TLS support of the new protocol
lib/kernel/src/net_kernel.erl
lib/kernel/src/dist_util.erl
lib/kernel/src/inet6_tcp_dist.erl
lib/kernel/src/inet_tcp_dist.erl
lib/kernel/src/net_kernel.erl
lib/ssl/src/inet_tls_dist.erl
lib/ssl/src/ssl_tls_dist_proxy.erl
lib/kernel/doc/src/net_kernel.xml
It seems to me that we could roll it out in two phases:
1. Implement 1,2,3 in the first phase. This would allow to deploy a
version with EPMD supporting old and new protocols.
2. Implement changes to net_kernel and transports (4,5) to be able to talk
to new EPMDs
If agreed, I can break up my patch in two, corresponding to this plan.
What do you think?
Serge
Post by unknownPost by unknownhttps://github.com/erlang/otp/pull/121
Has it been reviewed?
Thanks!
Serge
Hi Serge,
We have discussed your contribution at the OTP technical board and here
Inclusion in OTP
-----------------------
Your patch touches many source files which are central for the Erlang
distribution and we feel it would require quite a lot of work from our side
as well to assure that nothing in the current distribution gets broken or
incompatible. It would also require quite deep involvement from us giving
more detailed feedback regarding interfaces, semantics etc. We can not
prioritize this at the moment so we will not include your contribution now
(which means REJECT for now).
But we really want to add this kind of functionality and want to encourage
you (and others) to continue the good work in this area for possible
inclusion in OTP at a later stage. See more info and thoughts below.
General thoughts
-----------------------
Enhancing/extending the Erlang distribution to support heterogenous
communication protocols is something we have discussed many times and it is
something that we really would like to have.
1) We would like a plugin architecture where different transport protocols
can be easily plugged in without having to change the basic Erlang
distribution or the VM.
2) It should be possible to run different protocols towards different nodes
3) should be possible to have different timeouts (NET_TICK_TIME) per
connection
4) interesting protocols would be UDS, TCP, TLS over TCP, SCTP, ...
5) it should be possible to look up nodes in different ways and maybe not
only via contacting epmd on the host name which is part of the node name.
I.e having the hostname as part of the node name is not always what you
want, depending on protocol to use.
....
Your specific contribution
-----------------------------------
In your pull request you have implemented fully or partly 1,2, 4, and we
see it as a good start.
A general comment is that the diffs are showing differences on many places
in the code where you actually haven't changed anything except tabs, line
breaks and spaces. This makes it more difficult to sort out what you have
changed and added. We would like a clean diff without unnecessary touches
lines which has no relevance for the new function. If a formatting change
is needed or suggested it should be a separate commit/pull request.
Regarding details in your code we have not had the time to take a deeper
look but from a quick glance it looks good. The questionmarks are more
about how to introduce this new functionality in a safe way and what more
is needed that you have not addressed so far.
Suggested way forward
------------------------------
I think that one approach that would make it easier to include this kind
of functionality soon would be to make minimal changes to the existing
distribution code and only add or change so that it is possible to support
heterogenous protocols with code that initially can live as a separate
extension that don't need to be part of the OTP distro.
In this way the functionality can be further developed and used together
with an ordinary OTP distro (unpatched) and can easily be included later
(hopefully within a year).
Maybe you think that you already have done those minimal changes, but I
suggest you take another look and see if it is possible to find a way to
make minimal or now changes to net_kernel and epmd etc. which we could
include and then have the rest as a separate "application"/extension
project at github that can be included later.
Maybe let the old distribution work as before and let the user choose
"new_distro" in some way and will only then get to run the new code which
allows heterogenous protocols another new EPMD or an EPMD written in Erlang
or allow for plugins of different EPMD implementations that can run in
parallell. The communication towards EPMD maybe needs to be secure as well
and then it might be much easier to implement in Erlang.
I (we) are more than willing to discuss more about how to find this
minimal patch which we think we can "safely" include and which makes it
possible to develop and use the rest as pure extensions to OTP in a first
step.
Regards Kenneth, Erlang/OTP Ericsson
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-patches/attachments/20140109/1894bd3c/attachment-0001.html>