The sampling probability is now being computed using an estimate of the table row count
This could led to too high probabilities (to large samples) if the estimate is not accurate.
To avoid potential problems with large samples we've added a LIMIT to the sampling queries.
Keep current production behavior of ignoreing errors when computing this stat and returning -1.
This is done as to no introduce any instability in production at the moment.
Remove usage of PhasedExecution
This achives better query execution granularity and
removes questionable usage of shared results object.
It introduces a couple of behavior changes:
* estimatedFeatureCount desn't ignore errors now
* sample always uses estimatedFeatureCount,even if the actual count is also computed.
Javier Goizueta