OK I know this has been long delayed but I've finished some work on the
above. The coster is actually doing a fairly good job. I only recieved
one submission from someone with data that replicated the problem, and
was myself hard pressed to replicate the situation. It's more-or-less a
fencepost error. I don't have the expertise to figure out how to make
the coster more determinate in these types of situations. However as
some suggested the practice of storing actual run data from query plans
(esp. when using precompiled and/or stored queries) would probably help
eliminate these byt adding another weight factor (IE last time we did
this it took X amount of time, and we estimated Y, so lets try it this
way instead).
Unfortunately I'm a bit too pressed for time looking for a job to
continue pursuing this research any further.
Michael Loftis
