Do ROWID Index Row Entry Columns Impact Index Block Splits ? December 20, 2007Posted by Richard Foote in Oracle Indexes, Richard's Musings, Index Internals, Concatenated Indexes, Index Block Splits, ROWID.
Based on a great question by Alberto Dell’Era in my “Differences Between Unique/Non-Unique” blog entry (comment 9), I thought it might be a useful exercise to show how I go about confirming my understanding of a specific concept by trying to develop a little test case or demo that can illustrate the concept. My “magic incarnation” if you like
The basic question was does the ROWID that constitutes an additional column in a Non-Unique index determine whether a particular row entry is the maximum or equivalent entry or not. Because by implication, this can determine and influence whether Oracle performs the generally preferred 90-10 splits rather than 50-50 block splits for indexed column values that at least equal the maximum value.
The answer is yes because the ROWID column is just another column in the index row entry and is simply treated the same. But how to actually “illustrate” and show this ?
I needed a way therefore to insert a ROWID that was always going to be the maximum ROWID value for a Non-Unique index. Then insert a whole bunch of subsequent ROWIDs of a lesser value than the maximum and inspect via index statistics whether the type of block splits changed from 90-10 to 50-50 block splits. Remember with the Object Number being equal (if it’s there at all), the next significant portion of the ROWID is the Relative File Number.
The plan was (reasonably) simple. Create a tablespace with one data file and fill it with something. Then add a second data file and use this to store the start of my table of interest (and of course create the index). This will create a whole bunch of rows with ROWIDs of a higher Relative File Number than those in the first data file. Then drop the first table and ensure the second table uses the free space created in the first data file. That way, a whole bunch of ROWIDs can be created that are less than existing ROWIDs because it would be using ROWIDs from the first data file, which has a lesser Relative File Number.
It’s the usual process I go through with these things. Find something that’s of interest, have some idea on how I think things work, come up with plans or strategies that will illustrate whether or not what I think is true (ensuring that somewhere in the process I include at least one reference to David Bowie . I can then later take the initial strategies and expand them for all applicable database options and features. Then see if anything changes between database versions and platforms.
Hopefully this demo shows you how I went about proving this: Do ROWID Index Row Entry Columns Impact Index Block Splits Demo.
The benefit of then showing these demos is that others can see exactly how I came to a conclusion, potentially try them out for oneself and perhaps see holes or flaws or shortfalls in the strategy or expand or tailor them for individual requirements or environments.