Any thoughts on stats validation?
September 7, 2011 4 Comments
There is an application running on 22.214.171.124 and this application’s approach to object statistics is that they are gathered along with each quarterly release of code and do not change outside of the release schedule.
There are a number of other additional issues with some elements of this particular implementation, discussion of which are outside the scope of this.
Now, as part of the upgrade to 11.2, it’s highly likely that we will move to the default built-in stat proc, albeit gathering in a weekend window not daily.
One of the biggest advantages of the current situation is that the stats are tested along with the release and released with the release (of course, plans can still change depite stats being fixed – even because they are fixed due to time ticking ever onwards, sequences increasing, etc).
And so, as part of a possibly non-negotiable comfort blanket to move to the default 11g stats job, there is a management requirement to have some sort of validation routine – i.e. a set of sql statements broadly representing important or problem areas of the application – that will be run before/after and compared on some sort of metric and judgement passed on whether the stats are “good”.
There are a number of nifty features like pending stats that we can use here but my initial thoughts on this was that this sounded like a fairly easy job for SQL Tuning Sets (STS) and the SQL Performance Analyzer (SPA).
SPA requires the additional Real Application Testing (RAT) license and, now that our licensing has been negotiated, this isn’t something we’ve got.
So… hand-rolling a validation routine.
I would see SQL Tuning Sets with the sql statements and bind sets as still part of a bespoke solution
(STS are available with the Tuning Pack license which is on top of EE both of which we have).
I could picture a simple approach that would execute these statements in parallel sessions probably via DBMS_SQL, fetch all the rows, record the session stats before/after, pick a metric e.g. buffer gets, elapsed time, or some sort of ratio between those sorts of things and then and do some sort of metric comparison – yes, that run was broadly the same or better, or at least not a catastrophic degradation – these stats are ok.
(Of course, being able to picture a possible approach does not mean that I advocate it)
But I just wondered whether this is something anyone else has done?
I’d be very interested in hearing more if you have.
I’ve tried to reword this to avoid some confusion.
Alex – I’ve deleted your comments and my replies because a) they reduce the chance of getting the input the article was aiming for b) they were based on nested recursive misunderstandings and c) much of what you were saying is preaching to the choir.