For edge case weights like { 5, 1, 1 } we now produce { a, a, b, a, c, a, a }
sequence instead of { c, b, a, a, a, a, a } produced previously.
Algorithm is as follows: on each peer selection we increase current_weight
of each eligible peer by its weight, select peer with greatest current_weight
and reduce its current_weight by total number of weight points distributed
among peers.
In case of { 5, 1, 1 } weights this gives the following sequence of
current_weight's:
a b c
0 0 0 (initial state)
5 1 1 (a selected)
-2 1 1
3 2 2 (a selected)
-4 2 2
1 3 3 (b selected)
1 -4 3
6 -3 4 (a selected)
-1 -3 4
4 -2 5 (c selected)
4 -2 -2
9 -1 -1 (a selected)
2 -1 -1
7 0 0 (a selected)
0 0 0
To preserve weight reduction in case of failures the effective_weight
variable was introduced, which usually matches peer's weight, but is
reduced temporarily on peer failures.
This change also fixes loop with backup servers and proxy_next_upstream
http_404 (ticket #47), and skipping alive upstreams in some cases if there
are multiple dead ones (ticket #64).
Previously nginx used to mark backend again as live as soon as fail_timeout
passes (10s by default) since last failure. On the other hand, detecting
dead backend takes up to 60s (proxy_connect_timeout) in typical situation
"backend is down and doesn't respond to any packets". This resulted in
suboptimal behaviour in the above situation (up to 23% of requests were
directed to dead backend with default settings).
More detailed description of the problem may be found here (in Russian):
http://mailman.nginx.org/pipermail/nginx-ru/2011-August/042172.html
Fix is to only allow one request after fail_timeout passes, and
mark backend as "live" only if this request succeeds.
Note that with new code backend will not be marked "live" unless "check"
request is completed, and this may take a while in some specific workloads
(e.g. streaming). This is believed to be acceptable.
