There's only so much duct-tape you can throw at a problem before it becomes unmanageable. The scale, complexity and reliability of AWS indicates that they are indeed running a tight (and tidy) ship.

It's a mix. The code is good enough to get things running, but it still needs full time oncall who actually are constantly putting out little fires.

I have heard whispers of this, and the mention of common on call hours seems to lend it some credibility

Yes, seriously legacy software (still a lot of Perl in their production web code, including ancient Perl template engines) and the on-calls are truly bad. AWS is more modern, but still not entirely awesome from what I've heard.

On the plus side, hiring people out of there after a year is usually pretty easy.

What are the typical on-calls?

You are paged most nights and often more than once a night during your rotation and you are on rotation a lot.