Sql-server – the significance of root and page level fragmentation in a clustered b-tree index

It sounds like the rebuild of the index is a red herring, especially given it was non-clustered. The rebuild is more likely to be invalidating a bad query plan in cache, and on re-run the good query plan is generated and used again.

Captue the query plan when the performance is bad and then after the rebuild and compare them, I suspect you will see a different plan.

That would be my first step - e.g. demonstrate what changes to the plan due to the index rebuild, or prove there is no change on the plan.

FILLFACTOR only applies when you build or rebuild an index, not during normal operation. Normal operations always try to fill the pages to 100%.

If you insert a row that has a variable width, then update the row to be longer, that row will no longer fit on the page if there isn't enough extra space to store the after-image on the same page. If there isn't enough space, this will cause a page split, which is the process that creates the necessary space.

What's a bit misleading about page splits is that there are "good splits" and "bad splits," even though the performance counter counts all of them.

• A good split is when new rows are added to the end of the index, like what happens when you do your initial batch of INSERTs. The new row doesn't fit on the last page, so the storage engine must allocate a new page and logically hook it up to the last index page. The new page will probably exist physically after the old-last index page.

• A bad split is when a page must be inserted in the middle of an index: the new page is attached to the index structure logically on both sides, but may (likely) not exist in contiguous physical order to those pages.

Fragmentation is the discrepancy between the logical and physical order, and for the most part, only the bad type of page splits cause fragmentation.

Because you're inflating the size of already-existing rows, this causes the bad type of page split, which is why you're seeing high fragmentation numbers.

It's unclear what your exact process is, and how many rows are ultimately going to end up in this table. If it's something like a one-time population, do the full population process, then turn around and rebuild the clustered index with 100% FILLFACTOR.

If this process happens continuously, you could "pre-allocate" the space by doing the following: add a dummy variable-width column to the table, populate it to the max length on INSERT, and then on the first UPDATE, set the dummy value to NULL while updating the real value. This method will probably add overhead to the logging mechanism because of all the data values flying around.

In general, though, you only need to concern yourself with fragmentation when all of these factors are true:

• The index is large enough that randomly reading the pages (extents) is too slow (choose your own definition of what "too slow" means)
• The index will be scanned
• The index pages (extents) will be physically read from disk (i.e., they are not already in the buffer pool)
• The storage subsystem is poor at handling random reads (i.e., spindle-based, and the bytes won't come out of cache)

I would recommend, however, putting in place a solid index maintenance solution to keep things more or less in check. You can search this site for recommendations with regards to that.