[gaius]

When Altavista launched, it was an impressive showcase of the DEC Alpha's power. Intel only became usable for serious servers (with the exception of exotic stuff like Sequents), as did Linux, years later. Google had the good fortune to be in the right place at the right time, when Lintel became a commodity in the datacentre. 5 years earlier, they'd have been on Sun probably.

[paulsutter]

Altavista launched in 1995 and Google began as a research project in 1996. At my own startup, in 1996, we used Intel because with Sun servers you paid an extreme markup for unnecessary reliability.

I was VP of Engineering at Altavista in 2000, and I started the project to move to Linux. It wasn't easy because search engineering was populated by Alpha fans who were unswayed by the 10x cost advantage.

As late as 2001, I sat in multiple focus groups where all the enterprise customers said Linux was not yet ready for the datacenter. IBM's penguin campaigns were just beginning at that time.

Google's large scale use of Linux was groundbreaking when they launched in 1998.

[minimax]

Initial versions of Google at Stanford ran on a mix of both Linux and Solaris.

http://infolab.stanford.edu/~backrub/google.html

[paulsutter]

I'll bet you a dollar that Larry and Sergei never actually bought a Sun server to run the search engine, but rather that they may have used some free resource available to them at Stanford.

[anamax]

> When Altavista launched, it was an impressive showcase of the DEC Alpha's power.

Altavista was started by Paul Flaherty. One of his jobs was to find some way to showcase Alphas.

[aswanson]

Do you think that would have affected their prospects for success?

[gaius]

It would have affected their cost base certainly, and probably their entire datacentre strategy. With SPARC kit, you wouldn't build assuming that machines will often fail and simply be swapped out, for example, something that Google is famous for.

[anamax]

> With SPARC kit, you wouldn't build assuming that machines will often fail

In the late 90s SPARCs did fail. Yes, they were more reliable than commodity x86 boxes, but they failed often enough that it was an issue if you had 100 or so, and search engines hit that level very quickly.

[gaius]

Right, but look at what Google do, their boxes are basically disposable. Why invest in dual-redundant-hotswappable-everything boxes when you just throw the entire thing away if any bit of it breaks, 'cos it's cheaper to replace it than to even try to repair it in-place.

[anamax]

> Right, but look at what Google do, their boxes are basically disposable

We're talking matter of degree.

The claim was that building a search engine out of 90s sparcs meant that you didn't have to worry about things dying.

That claim is not true - reasonable search engines of that era required enough machines that the failure rate of 90s sparcs, while better than x86 of the time, was enough to require folks to handle frequent failures.

It's reasonable to argue that the cost/benefit tradeoff of sparc's extra reliability vs x85 wasn't worth it for those companies, but that's a different argument.