named

[Richard Pieri]

On Dec 2, 2010, at 11:47 AM, Derek Martin wrote:
> 
> You're missing the point.  Your statement is true, as far as it goes;
> but what is an end-user going to notice more?  The average case of a
> locally cached DNS lookup taking a few milliseconds less time, or the
> edge case of uncached or expired lookups taking 10s longer?  Edge
> cases do matter if they happen enough, and perception is reality. 

That depends on how frequently my users fall into the edge cases.  There are things that I can do about that on a case-by-case basis if it becomes a significant issue.


> It doesn't matter if it's geographically local... all that matters is
> network topology.  It's rare that an ISP's DNS servers are not
> topologically closer to their end users than... well, much of anything
> else.  It does happen, but it's an edge case that can be identified
> by CDNs and worked around in other ways, once traffic is actually
> being served to those IPs.

So, what happens when the shortest link breaks?  Say a router somewhere gets overloaded or fails; backhoe cuts a fibre.  The shortest path is no longer the best path, maybe not even a valid path, and traffic gets rerouted -- except for the CDN which still thinks that the shortest path is the best path.  It isn't a simple problem, and relying on a simple solution starts to fall apart when faults occur.  But this is getting into the reams of why DNS roulette for failover is bad.


> Any technical arguments against it do nothing to change the fact that
> its use is widespread.

So are Microsoft Windows and Oracle.  Doesn't make them good.

--Rich P.