It's interesting that they're doing a spatial database without using a 2D spatial data structure. MySQL and PostGres both have data structures for this. But maybe they don't have that many geofences per city.

They have an N readers, one writer problem. It's possible to do that without blocking the readers. There was a YC article on a lockless solution to that yesterday. This is an easier problem. Each city's geofences can be updated by making a copy, updating the copy, and atomically replacing the single reference to the copy. Any searches still running on the old copy continue to run on the old copy. Eventually, the GC deletes the old copy.