I started the series here: http://www.sqlskills.com/blogs/kimberly/post/Spring-cleaning-your-indexes-Part-I.aspx and I want to continue with Part II today by clarifying some great questions/comments that have come up on the series. In Part III, I'll give you a few more ways to get rid of (or consolidate) indexes. And, I think there's still a bit more that Paul and I will investigate further (wrt to operational stats) but, I want to address a few comments and a few interesting things that both Paul and I have found.

In the Part I post, I talked about using sys.dm_db_index_usage_stats to see if there are any indexes that just aren't being used at all… A few comments asked why I didn't use operational_stats instead. To address that first, there are a few key differences:

  • dm_db_index_operational_stats is persisted only as long as an object is in cache (however, it's not cleared when objects are forced out of cache with DBCC DROPCLEANBUFFERS). If you want to clear ALL DMVs for a specific database, then a relatively easy way to do this (IN TESTING) is to take the database offline and then immediately bring it online again.
    • ALTER DATABASE <dbname> SET OFFLINE
    • ALTER DATABASE <dbname> SET ONLINE

      • NOTE: If there are any suspect files, you will NOT be able to bring your database back online without FIRST taking all suspect files OFFLINE. And, if you take a FILE offline then it's even more important to know that THERE IS NO WAY TO BRING A FILE ONLINE without restoring it from backups. So, it's VERY important to understand that OFFLINE/ONLINE for a database is really easy IF AND ONLY IF there are no other problems with the DB. You really need to resolve those problems first (or at least know that you're going to need to resolve those problems later through backup/restore) before you take a database offline.

  • dm_db_index_operational_stats is (from BOL) neither persistent nor transactionally consistent. This means you cannot use these counters to determine whether an index has been used or not, or when the index was last used. For information about this, see sys.dm_db_index_usage_stats.

Having said that though, none of these are really any guarantee of perfect information. And, they're not meant to be. I look at these DMVs as being a quick and easy way to get some relatively descent insight into what is or is not happening in my environement. However, even though the sys.dm_db_index_operational_stats might give you insight that you have a problem it still doesn't give you good insight into exactly what that problem might be. For example, it *might* be splits that causes some of your wait times to increase (column: page_io_latch_wait_in_ms) but, it could be something else too (some other system issue).

The main point, you can use these to get insight into which tables have the biggest problems (i.e. the biggest waits) and where they might have a lot of splits (column:nonleaf_allocation_count) but, in all honesty, that's not a guarantee. In fact, the reason I said "might" is that pages that are allocated at the end of the leaf level STILL allocate a page and require an entry to be made in the next level up in the index. So, a lot of nonleaf_allocations COULD be for a perfectly unfragmented index. So, it still doesn't tell you how fragmented the objects are or what the REAL problem is (or even if it is a problem yet).

Basically, these just tell you where you have the most activity and give you a starting point for problem solving. But, none of these (usage or operational) really tell you how to solve the problem. However, sometimes even knowing where to start IS the problem in and of itself. So, I'm not against these DMVs and I really do think you can get some good insight from them. Just use them as a tool to help focus your investigations. Use better tools like sys.dm_db_index_physical_stats to really see if you have fragmentation and where it's the worst.

hth,
kt