I am certain the are normally writes going on; they do run Analytics on their homepage. However they get to defer, retry and play lots of eventually consistent tricks, worst case just swallow the exceptions. The fact they can make the service _seem_ to the end user as fully working whilst being unable to write is a major factor in achieving their world beating reliability.

Sure, but they can have several copies of the index per datacenter, retry your query multiple times posibly even in a diffrent datacenter. New code and even updated indexes can be tried and then fall back to yesterday's version.