Josh Berkus <josh@agliodbs.com> writes:
> Well, for the straight test on event_time, the rows in the table should
> have been more-or-less physically already in order. This table is
> most-insert, and event_time mostly corresponds to current server time.
That would result in a very long initial run (probably all or most of
the data), which fits with your observation so far. But we don't know
what happens after that.
BTW, are you paying attention to vmstat or similar measure of the amount
of I/O and CPU work getting done? vmstat data ought to make it possible
to estimate the size of those runs, even though the trace_sort logic
isn't telling us.
regards, tom lane