I don't and won't have evidence unfortunately, only theory. The theory I am applying is that provided the values are contigious, the mapping used in the btree algorithm can use this to calculate the physical location of the page more quickly than if not and that page seeking will slow down proportionately according to how often it cannot calculate the page location based on the value and has to go walkies looking for it. OK I kind of have evidence because I have subjectively noticed a performance improvement in operating on large tables where the values of the PK are contiguous relative to those where they werent. In fact where PK values were contiguous Postgres crunched through 4 mil records in seconds but slept soundly for same record count but non contiguousPK values. Please don't ask me to submit test results -- I have a system to build!