The PowerDNS Recursor shuffles answers randomly, so no single A record gets overloaded. This logic also took care not to shuffle a CNAME record until after the name it points to, because we theorized this would upset some resolvers.
Our logic however assumed all the CNAMEs would initially be at the front of the packet. We'd start our shuffling after skipping all the CNAMEs up front. It now turns out that sometimes we end up with a 'CNAME A CNAME A' packet to shuffle.
This would happily shuffle the last three records. With this PR, we put the CNAMEs up front explicitly before commencing the shuffle. Closes #4339.
Still to be investigated: why didn't this bite us before?
projects / pdns / commit