Some questions about retry in ewma balancer

[spacewander]

After reading https://github.com/kubernetes/ingress-nginx/blob/master/rootfs/etc/nginx/lua/balancer/ewma.lua. I have some questions:

1. During retry, the tried endpoints still be counted when picking a new upstream. Will a request hit a bad endpoint which has lowest score twice?
2. I notice only the first upstream will be counted during retry:

   ingress-nginx/rootfs/etc/nginx/lua/balancer/ewma.lua

   Lines 183 to 186 in 7a02609

   	local response_time = tonumber(split.get_first_value(ngx.var.upstream_response_time)) or 0
   	local connect_time = tonumber(split.get_first_value(ngx.var.upstream_connect_time)) or 0
   	local rtt = connect_time + response_time
   	local upstream = split.get_first_value(ngx.var.upstream_addr)

   Should not we count the latest/last upstream? A tried upstream will be counted multiple times, is this intentional?

[aledbf]

/assign @ElvinEfendi

[ElvinEfendi]

@spacewander yes, there's a chance that tried and failed endpoint gets picked again. Yes, it is also an issue that we only count the first one in every retry. Changing that to last one makes sense to me, effectively it means we count every try. It almost feels like the algorithm is designed with the assumption that NGINX prepends every next try to those upstream_ variables (whereas in fact NGINX appends).

Are you interested in sending a patch? I'd be happy to further assist.