[Foundation-l] [Internal-l] Relocation Announcement

[Gregory Maxwell]

On 9/24/07, Michael Bimmler <mbimmler at gmail.com> wrote:
> Following up on this (as I was asked this recently by a journalist as
> well): Is there a reason why we don't "decentralise" our servers more,
> be it inside the US and/or on the global scale.

There are many reasons things are the way they are. Some of the
reasons are technical:

Some parts of the site are technically hard to distribute: databases
synchronization doesn't work fantastically across long distances
especially if you can't be darn sure the link will not be severed.
Distribution can cause a reduction in reliability and lead to complex
failure modes which are difficult to diagnose and avoid.

Fortunately, more than 99.9% of all the requests are requests for text
and images which have not changed and are as a result cachable and
trivially distributable.  This part of the site is already
distributed, and it's fairly easy to argue that it's the only part
which needs really to be distributed for practical reasons.

Wikimedia has remote caches in the Netherlands and Korea.

Why are there not more remote caches?  There are several factors, two
of the largest ones:

(1) System maintenance considerations. (It's harder and more expensive
to maintain many small sites than a few larger ones).

(2) Bandwidth pricing.  The cost per megabit of connectivity goes down
which you buy in larger units.  If we split the traffic up into 20
parts we  could reasonably expect to pay 3-4x the amount per megabit
of peak capacity.

Of course there are a lot of other reasons, but they get fair too
technical for this list. :) I'm not saying that there isn't an
opportunity for some more cache sites (and, in fact, I've advocated
having a little more for a while) but distribution is no magic bullet
and such changes need to be driven by careful well informed
consideration.