This should (i expect) work with any number of backup servers, as long
as you only need 1 active.
Ryan O'Hara schreef op 30-1-2014 19:34:
On Thu, Jan 30, 2014 at 07:14:30PM +0100, PiBa-NL wrote:
Im not 100% sure but if i remember something i read correctly it was
like using a "stick on dst" stick-table.
That way the sticktable will make sure all traffic go's to a single
server, and only when it fails another server will be put in the
sticktable that will only have 1 entry.
Yes. That sounds accurate.
You might want to test what happens when haproxy configuration is
reloaded.. But if you configure 'peers' the new haproxy process
should still have the same 'active' backend..
p.s. That is if im not mixing stuff up...
This blog has something very close to what I'd like to deploy:
http://blog.exceliance.fr/2014/01/17/emulating-activepassing-application-clustering-with-haproxy/
The only difference is that I'd like to have more than just one
backup. I'll try to find some time to experiment in the next few days.
Thanks.
Ryan
Ryan O'Hara schreef op 30-1-2014 17:42:
I'd like to define a proxy (tcp mode) that has multiple backend
servers yet only uses one at a time. In other words, traffic comes
into the frontend and is redirected to one backend server. Should that
server fail, another is chosen.
I realize this might be an odd thing to do with haproxy, and if you're
thinking that simple VIP failover (ie. keepalived) is better suited
for this, you are correct. Long story.
I've gotten fairly close to achieving this behavior by having all my
backend servers declared 'backup' and not using 'allbackups'. The only
caveat is that these "backup" servers have a preference based on the
order they are defined. Say my severs are defined in the backend like
this:
server foo-01 ... backup
server foo-02 ... backup
server foo-03 ... backup
If foo-01 is up, all traffic will go to it. When foo-0t is down, all
traffic will go to foo-02. When foo-01 comes back online, traffic goes
back to foo-01. Ideally the backend servers would change only when it
failed. Beside, this solution is rather ugly.
Is there a better way?
Ryan
