We have shown that client-side instability in DNS name lookups is widespread and relatively common. The failure cases degrade average lookup time and increase the ``tail'' of response times. We show that these failures appear to be caused by temporary nameserver overload, and are largely uncorrelated across multiple sites. Through analysis of live traffic, we show that a simple peering system reduces response times and improves reliability.
Using these observations, we develop a lightweight name lookup service, CoDNS, that uses peers at remote sites to provide cooperative lookups during failures. CoDNS operates in conjunction with local DNS nameservers, allowing incrementally deployment without significant resource consumption. We show that this system generates low overhead, cuts average response time by half or more, and increases DNS service availability.