The implementation of random cursors on tiered tables – i.e., curtiered_next_random() – can be pretty skewed. We should come up with a better way to select random items.

In the current implementation we select a tier at random and then select a random item from that tier.  This will produce skewed results in two likely usage scenarios:

• Tiers have substantially different numbers if items
• There are many items replicated across tiers

In LSM we choose a random chunk, weighting by the number of items in each chunk. But with tiered cursors we don't know how many items are in each of the tiers.  Perhaps we could use the number of leaf pages in each tree/tier as a proxy.