I've been thinking about a new access method that would allow for more efficient changes to be made to out-of-cache indexes. The basic idea is that the access method would consist of two constituent tables. One table for buffering up recent changes (the ingest table) another with all other data (the stable table).

Data would be moved in the background to the stable table, which could be done either via a streaming mechanism or via batch updates.

The ingest table can be small (remain in cache). Changes to the stable table are made periodically - ameliorating the cost of distributing content amongst the stable tree.

A cursor on such a table would look for content in both the ingest and constituent tables.