[kartan]

I will say that if the system scales well horizontally the important metric is how many resources you need to achieve that amount of transactions.

If they used 70 4-core m3.xlarge that means that they can attend 114 transactions/second per 4-core server. And that looks slow without more data about what are the requirements for the login process.

But if you look at the peak of 150 requests per second, then probably it will cost more money to improve the performance than to actually just add servers.

At work we have over 1 billion users. But as it scales horizontally the relevant number is how many memory/disk database instances do you have. The same applies to the application layer for the number of transactions. In fact the development servers are slower than the live ones even that they have just tens of thousands of users. For our amount of users/transactions tuning for performance saves lots of money. So it makes sense to improve the code on that measure.