Generally, top scientists/engineers are not motivated by money, which means given an offer from a typical ("glue a product together") start-up and Google they'd choose Google, unless they're looking for the "start-up experience".
If you want top technical talent (who will always be able to find fulfilling and challenging work at the whatever the current Google-like company is, e.g. SGI in early 90s, Netscape in mid 90s, etc), you have to give them top technical challenges (which is how these companies have themselves been able to compete for talent when they were start-ups).
In the next 6 to 18 months, I am going to have to start swapping out the big building blocks with parts meant to scale. At that point there should be plenty of room for the systems engineers to join me.
Be careful, however, with treating scale as equivalent of performance. You have to design for scalability, not optimize for it: given the same algorithm, sorting a list of items in python might take 50 times as long as doing it in C; yet using an n^2 vs. n log(n) algorithm a 1 000 000 item list would take 50000 times as long to sort.
I'm a ACM ICPC (programming contest) finalist, and have competed in the Google Codejam, so I know about big-oh.
Anyway, who implements a sorting algorithm in pure Python? I use the builtin sort() function, which is implemented in C (making callbacks into python for comparisons for some types). It uses TimSort - an approach designed to minimise the number of comparisons. http://svn.python.org/projects/python/trunk/Objects/listsort...
I thought you were going to point out that I should ensure that my system is modular (stateless, partitioned, etc) and I can spread the load across several machines - I can.
You are right that I could make a screaming fast bubble sort and still end up with major issues. Fortunately I already have had that epiphany while solving ACM practice problems.
I appreciate your advice, there are far too many who haven't yet understood it.
The sort example was merely a metaphor for design for scalability vs. optimization for performance. Perhaps better examples would have been dividing the server side portion of your application into asynchronously invoked services is design for scalability. Rewriting some of these services in C or OCaml, switching from JSON to Protocol Buffers, switching from an HTTP server and a layer 7 load balancer to a custom non-blocking server and ZooKeeper (for cluster membership) are optimizations for performance/stability/cost.
We use Python, Twisted, Django, Matplotlib, JQuery, Movable Type, Postgresql all for this reason. Even our firmware is built as much as possible on a 3rd party neworking library and high level hardware building blocks.