We currently have a ~1.5 TB VLDB (SQL 2008 R2) that services both OLTP and DSS workloads pretty much on a 24x7x365 basis. For many years we have been updating statistics (full scan- 100% sample size) for this VLDB once a week on the weekend, which is currently taking up to 30 hours to complete.
Somewhat recently we have been experiencing intermitent issues while statistics are being updated, which I doubt is just a coincidence. I'd like to understand exactly why the process of updating statistics can cause these issues (timeouts/errors). My theory is that the optimizer is forced to choose an inferior execution plan while the needed statistics are in "limbo" (stuck between the "old" and the "new"), but that is again just a theory. I'm somewhat surprised that the "old" statistics couldn't continue to get used while the new/current statistics are being generated (like the process for rebuilding indexes online), but I don't know all the facts behind this mechanism yet so that may not even apply here.
I understand that we have the option of reducing the sample percentage/size for updating statistics, which is currently set at 100% (full scan). Reducing the sample percentage/size for updating statistics will reduce the total processing time, but it's also my understanding that doing so will leave the optimizer with less than optimal statistics for choosing the best execution plans. This seems to be a classic case of not being able to have one’s cake and eat it too.
So in a nutshell I'm looking to fully understand why the process of updating statistics can cause access issues and I'm also looking for best practices in general for updating statistics of such a VLDB. Thanks in advance.
Bill Thacker