Upstream (fair) with dead servers

[Kamil Gorlo]

2009/7/9 Igor Sysoev <is at rambler-co.ru>:
> On Thu, Jul 09, 2009 at 05:09:56PM +0200, Kamil Gorlo wrote:
>
>> 2009/7/9 Igor Sysoev <is at rambler-co.ru>:
>> > On Thu, Jul 09, 2009 at 04:47:43PM +0200, Kamil Gorlo wrote:
>> >
>> >> Hi,
>> >>
>> >> I have (possibly quite rare) situation where I have one upstream with
>> >> two servers and I know that in any particular moment only one of those
>> >> servers is running:
>> >>
>> >> upstream blah {
>> >>     server server1; // when working, server2 is dead
>> >>     server server2; // when working, server1 is dead
>> >> }
>> >>
>> >> So, what I need is that Nginx will choose working backend (in this
>> >> upstream), remembers it and choose it for every request till some
>> >> failure on this server occurs. Then it tries other server in upstream,
>> >> till it finds some working - and situation repeats. Is this possible?
>> >
>> > No, nginx will send request on both servers in round robin mode.
>> > If one of them will fails, then nginx will resend the failed request
>> > to another server and mark the server as dead for fail_timeout time.
>>
>> Hmm, what about situation when all servers are marked as dead? Nginx
>> will wait till fail_timeout expires for some server and in meantime
>> return error to client? Or will try dead servers?
>
> In this case nginx marks all servers as live and try again.
>
>> Also, is it hard to write such a "upstream-chooser" which will work as
>> I described?
>
> I'm going to rewrite upstream modules in 0.8.x (add fair, etc.) and
> probably add this functionality. But why do you want to send request
> to one only server leaving the other as spare one ?

Those servers are not HTTP servers, but memcached-like servers (Tokyo Tyrant).

In my configuration Tokyo Tyrant is in master-master taste - which
leads to problems with replication when clients are writing to more
than one master (but this is master-master not master-slave because I
want that, in case of failure, the other working server will cope with
writes). So to prevent problems I have simple (and maybe ugly;)
scripts which blocks interfaces at non-current-master machine.

>> >> Maybe upstream_fair module with option no_rr will work for me (I am not sure)?
>> >>
>> >> Cheers,
>> >> --
>> >> Kamil Gorlo
>
>
> --
> Igor Sysoev
> http://sysoev.rua/en/
>
>