From the V2.0 Release Notes:

Your driver may also support tag-aware reads. Instead of simply specifying slaveOk, you specify slaveOk with tags indicating which data-centers you want to read from. See your driver for details.

The behaviour I would like to see for optimal performance and reliability is this for reads:

• First, try any node in my local datacentre, with option to include/exclude master from this selection
• Then try master if local, if not already tried
• Then try any node from any in list of datacentres.

This would give me best performance by always reading from my local datacentre first, but would then give reliability by being able to use any available if my local datacentre had problems.

The description as it stands implies that to get the benefit of being able to read from another datacentre, I have to take the hit of a percentage of my reads always coming from that remote datacentre, even when I have a local node available.