fix(balancer) fix accidental ttl=0 switches

See #51

Some servers will report ttl=0 when they are on the very edge
of their own cached ttl. This should never happen for a record
that has a non-0 ttl.

This fix makes sure we require ttl=0 reported twice in a row before
we switch the loadbalancer.

Fixes #51

src/resty/dns/balancer.lua

@@ -358,7 +358,12 @@ function objHost:queryDns(cacheOnly)
     return true    -- exit, nothing changed
   end
 
-  if (newQuery[1] or empty).ttl == 0 then
+  -- To detect ttl = 0 we validate both the old and new record. This is done to ensure
+  -- we do not hit the edgecase of https://github.com/Kong/lua-resty-dns-client/issues/51
+  -- So if we get a ttl=0 twice in a row (the old one, and the new one), we update it. And
+  -- if the very first request ever reports ttl=0 (we assume we're not hitting the edgecase
+  -- in that case)
+  if (newQuery[1] or empty).ttl == 0 and ((oldQuery[1] or empty).ttl or 0) == 0 then
     -- ttl = 0 means we need to lookup on every request.
     -- To enable lookup on each request we 'abuse' a virtual SRV record. We set the ttl
     -- to `ttl0Interval` seconds, and set the `target` field to the hostname that needs