Replication/failover on OpenBSD?
From: Joachim Schipper (jDOTschipper_at_math.uu.nl)
Date: 09/29/05
- Next message: Max Bolingbroke: "Re: Hiding NATs with PF"
- Previous message: Greg Hennessy: "Re: Hiding NATs with PF"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: 28 Sep 2005 23:19:22 GMT
First off, two smallish points: I'm not quite an experienced OpenBSD
user (being mostly a Linux guy looking around, right now), and all of
this is basically for my amusement as it's major overkill in our
situation.
I've build and are currently maintaining a server for a small students'
organization I am a part of. We have little traffic, and can acquire
oldish machines for little or no cost via sponsors, the university, &c.
Thus, while performance is a personal point of pride, it's ultimately of
less concern than security (as in both 'not getting hacked' and 'not
losing data').
I'm thinking about replacing our current Linux server with an OpenBSD
box. This is mostly because of the relative ease - OpenBSD is stable,
secure, and a lot better documented than the hardened Linux From Scratch
derivative I installed. Not to mention less buggy, and pkg_add or
/usr/ports are remarkably painless if you're used to wading through
./configure options.
Since we have plenty of hardware, but none of it could be called
reliable, some sort of failover solution seems called for (of course,
RAID would take care of part of the problem, but that's just too easy to
be any fun). Last year saw some of our people bringing the server down
by damaging the power cable (still haven't got a lockable cabinet, damn
them!), a hard disk crashing after years of hard work under
not-so-perfect conditions, and at least one weekend of effective
downtime after I misconfigured the network. Over SSH. Oooops.
Application-level replication could take care of some of the work, but
is a pain to set up, and there'll always be someone who thinks up a neat
application that does not handle replication. Or naively installs/codes
a webapp.
We currently provide several services, some of which handle dynamic data
(SQL daemon, subversion repository, FTP and SAMBA for documents, and of
course, mail storage). Not everything can be replicated at the
application level (ftp, samba and the inevitable buggy webapp being most
troublesome), and it also makes configuration a lot more complex.
The canonical solution would be to use some form of shared storage,
either SCSI-ish stuff which we'll have big trouble laying our hands on
and even more trouble finding replacement parts (or even trained
members) for, or an NFS server on stored hardware which just moves the
problem. (And opens up quite a can of worms wrt locking.)
Rsync and the like are not good enough, as the system should be able to
go down without any data being lost (connections being lost is fine, but
everything that is reported as written to disk should be secure).
While researching this, I was quite happy to see drbd (www.drdb.org),
but it doesn't work on OpenBSD, obviously.
There's some sporadic stuff on the net about NBD for OpenBSD, which
looks neat at first glance until I discovered I couldn't find a client
for OpenBSD. If there's one, RAID-1 over NBD could work. (However, while
someone, according to Google, seems to have ported the nbd server from
nbd.soureforge.net to OpenBSD, the client still has a line #include
<linux/ioctl.h>).
Of course, in any case, hacking together some heartbeat-like
(www.linux-ha.org) is required, getting the firewall to understand all
this, and creating per-application failover scripts. ifstated, which
also turned up, may be of some value here.
Is there some option for OpenBSD that will do synchronous data
replication, meaning that I can, during the most heavy load, pull the
plug on the currently active server without losing any 'committed' data?
(Assuming well-behaved daemons, a working backup node, &c.; by
'committed' I mean that the daemon in question has told the client he
received it, or at least finished writing; please note that this
should work with PostgreSQL databases if at all possible!)
Ideally, two nodes should be allowed r/w access, but this is not
necessary (and of little practical value, as one node can easily handle
the load - and if not, failover will not help much anyway until we place
at least 3 boxes there).
If only I had time to try some kernel hacking...
Thank you all, and excuse me for the lengthy rambling,
Joachim
- Next message: Max Bolingbroke: "Re: Hiding NATs with PF"
- Previous message: Greg Hennessy: "Re: Hiding NATs with PF"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
