https://sourceforge.net/p/net-snmp/code/merge-requests/3/
I had quite a lot of fun on this project, and I'd like it to be merged
upstream. I know it's 71 commits, and may be easier to just read the final
'snmp_api.c'. I'm also guessing all the source owners are volunteers, who
are busy.
Anyway, it's been a few weeks, and I'm not complaining, but I'd appreciate
some comment on what you'd like to see improved to accept it before I
forget everything.
I did this to expose the asynchronous API of NetSNMP of old which I'm
accustomed to, and (predictably) have been able to realize some significant
improvements in monitoring delay, experimentally.
Things I learned / issues to spur discussion;
I used the new 'Capsule' API, which unfortunately makes it rely on Python
2.7 or later. In retrospect, it might have been better to make the Session
and PDU objects "custom types," so that they can be accessed by Python
users similarly to the C structures they approximate, rather than what I
wound up doing for the callback function - making a dict that approximates
the C structure.
I dug out placeholders for a grepped-out list of API functions. No doubt
some are internal and should be deleted, but it represents opportunities
for future work to expose more of the C API to Python users. I
prioritized what I needed, and did what I could in the week I had.
I've used this library since the CMU-SNMP days, and I noticed the C API was
updated in a way that reduced some functionality from old CMU->UCD->NET -
I'm referring to the abstraction layer for SNMP transports. It used to be
(or so I remember) there was a sockaddr_in on the PDU structure, so you
could open a session with no peername, and then set the target address on a
PDU-by-PDU basis (this is important, for example, in monitoring systems at
scale to control the number of open sockets and allocated socket buffer).
The new abstraction structure puts an opaque object there, and the only way
to get the same functionality is to create a 'indexed sockaddr' object
manually, guessing that's what the default (UDP) transport is going to
require. A little hinky, but it works. But even if you have a reference
to the non-default transport you're using, there isn't a transport API
function to resolve a new opaque object (say from a hostname or IP address)
on demand. I'd be happy to submit a C API extension for merging if there
is consensus - I think there should be a way to resolve an opaque object
onto a PDU from a {multi,single}-session pointer and peername. It should
throw an error on transports that don't support multiple endpoints (e.g.
TCP).
The netsnmp-python setup.py sets its version to 1.0a1. I think this
should simply be the netsnmp release version (and it should default to link
statically off --basedir=.. if possible, as most users build eggs which
should be as hermetic as possible). Any objection to changing one or both
of those?
We'd really like to get MIB modules loaded out of egg or wheel files as
well, talking about hermetic eggs, rather than having the runtime try to
find them in magic local directories. This allows us to build python
binaries that are fully self contained, and all changes (including mib
updates) follow our software release for same. To do this I think the C
API will have to change, so that we can insert an abstraction layer where
the Python package loader is used instead of system file open routines to
read the MIB files.
--
Network Tools & Automation
Twitter, Inc.
------------------------------------------------------------------------------
_______________________________________________
Net-snmp-coders mailing list
Net-snmp-coders@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/net-snmp-coders
