Slide 22 of 29
Valid times are unified to a single line.
The durations and times of different resolution, e.g. “January 1997” can be properly indexed by analogy with some quad-tree models: an object with a time-stamp of “Week 32, 1997” has its id appear in the index at an intermediate level - or several intermediate levels - whereas the finest resolutions “chronons” drop to the bottom as instants. Actual clustering on disc (time-close or space-close) depends on anticipated uses.
Open intervals (indefinite future, past) require hacks.
Do we need an index for transaction times ? Yes, it could be entirely distinct, especially since timestamps will be more uniform (e.g. 32-bit posix time) and thus more optimised algorithms could be used.
Modern Universal ORDBMSs allow configurers to add their own datatypes, and even their own indexing structures,e.g. Informix/Illustra DataBlades.
What about “version time”: higher level abstractions of transaction time ? Feasible, but not very useful ??
