Once up on a time, there was a database server with 500GB of data and a heavy read workload of dynamic queries. Data was updated frequently throughout the day and index tuning was a serious challenge. At the best of times, performance was dicey.
Then things went bad
Application performance plummeted. Lots of code changes had been released recently, data was growing rapidly, and the hardware wasn’t the absolute freshest. There was no single smoking gun– there were 20 smoking guns!
A team was formed of developers and IT staff to tackle the performance issue. Early in the process they reviewed maintenance on the database server. Someone asked about index fragmentation. The DBA manager said, “Of course we’re handling fragmentation!” But a few queries were run and some large, seriously fragmented indexes were discovered in production.
The DBA explained that fragmentation wasn’t the problem. She didn’t have automated index maintenance set up, but she periodically manually defragmented indexes that were more than 75% fragmented.
Bad, meet ugly
At this point the whole performance team flipped out. Trust disappeared. Managers squirmed. More managers were called in. The DBA tried to change the subject, but it was just too late. More than a week was wasted over Fragmentation-Gate. It was a huge, embarrassing distraction, and it solved nothing.
Here’s the deal– the DBA was right. Fragmentation wasn’t the root cause of the performance problem. The DBA was a super smart person and very talented at performance tuning, too! (And no, I’m not secretly talking about myself here– this is not the “royal she”.) But she made a strategic miscalculation: she should have set up occasional automated index maintenance to align with her team’s normal practices and standards.
Why you need automated index maintenance
When performance gets bad, one of the very first things people look at is whether systems involved are configured according to best practices. If you’re not following a best practice, you need to have a really good reason for it.
Regular index maintenance still has a lot of merit: even in Shangri-La, where your data all fits into memory and your storage system is a rockstar with random IO, index maintenance can help make sure that you don’t have a lot of empty space wasting loads of memory.
It’s still a good idea to automate index maintenance. Absolutely don’t go too crazy with it– monitor the runtime and IO use and run it only at low volume times to make sure it helps more than it hurts. Be careful, but don’t skip it.
How much downtime can you spare?
Before you implement index maintenance, find out how much time tables can be offline in each of your databases. Then, figure out what operations you want to use.
To Rebuild or Reorganize: That is the Question
First off: ‘Reorganize’ and ‘Rebuild’ are two different operations that each reduce fragmentation in an index. They work differently toward the same end. You don’t need to run both against the same index. (I sometimes find that people are doing both against every index in a maintenance plan. That’s just double the work and NOT double the fun.)
Rebuild: An index ‘rebuild’ creates a fresh, sparkling new structure for the index. If the index is disabled, rebuilding brings it back to life. You can apply a new fillfactor when you rebuild an index. If you cancel a rebuild operation midway, it must roll back (and if it’s being done offline, that can take a while).
Reorganize: This option is more lightweight. It runs through the leaf level of the index, and as it goes it fixes physical ordering of pages and also compacts pages to apply any previously set fillfactor settings. This operation is always online, and if you cancel it then it’s able to just stop where it is (it doesn’t have a giant operation to rollback).
Factors to consider:
- Standard Edition rebuilds ain’t awesome. If you’ve got SQL Server Standard Edition, index rebuilds are always an offline operation. Bad news: they’re also single-threaded. (Ouch!)
- Enterprise Edition rebuilds have gotchas. With SQL Server Enterprise Edition, you can specify an online rebuild — unless the index contains large object types. (This restriction is relaxed somewhat in SQL Server 2012). You can also use parallelism when creating or rebuilding an index– and that can save a whole lot of time. Even with an online rebuild, a schema modification lock (SCH-M) is needed at the time the fresh new index is put in place. This is an exclusive lock and in highly concurrent environments, getting it can be a big (blocking) problem.
- Rebuilding partitioned tables is especially tricky. You can rebuild an entire partitioned index online– but nobody really wants to do that because they’re huge! The whole idea behind horizontal partitioning is to break data into more manageable chunks, right? Unfortunately, partition level rebuilds are offline until SQL Server 2014.
- Reorganizing can be pretty cool. ‘Reorganizing’ an index is always an online op, no matter what edition of SQL Server you’re using. It doesn’t require a schema mod lock, so it can provide better concurrency. Reorganizing only defragments the leaf level of the index. On large tables it can take longer than a rebuild would take, too. But as I said above, it’s nice that you can reorganize for a while and then stop without facing a massive rollback.
SQL SERVER 2014: WAIT_AT_LOW_PRIORITY, MAX_DURATION, AND ABORT_AFTER_WAIT
I’m really excited about new index rebuild options that are shipping in SQL Server 2014. Check *this* out:
ALTER INDEX OhSoFragmented ON dbo.MyTable REBUILD WITH ( ONLINE = ON (WAIT_AT_LOW_PRIORITY (MAX_DURATION= 5, ABORT_AFTER_WAIT=BLOCKERS) ) );
So we’ve got new tools for those concurrency problems I was talking about with online rebuilds. Now, we can say how long we’re willing to wait to get that schema modification lock (in minutes), and if we have to keep waiting what should happen. (Options: kill those who are blocking us, kill ourself, or do nothing.) Obviously there are some situations where just killing the blockers could be a terrible problem, but I’m interested to experiment with this.
You Didn’t Answer the Question: Do I Use Rebuild or Reorganize?
Yeah, I totally dodged that question, didn’t I?
If you have a regularly scheduled downtime every weekend, you’re probably fine with straight up index rebuilds, even if you have Standard Edition. Single threaded offline index maintenance may not be the hottest thing in the world, but hey, if you’ve got time for it then embrace the simplicity.
If you have Enterprise Edition, embrace parallel index rebuilds– and use the ONLINE option for indexes that allow it if people need to access the database during your maintenance window.
If you have database mirroring or AlwaysOn Availability Groups, tread with caution– particularly with rebuilds. It’s easy to generate a ton of IO with index maintenance, and it could mean putting your secondaries or mirror so far behind that they can’t catch up.
Maintenance plans or custom scripts?
You can go the easy way and use SQL Server Maintenance Plans, but unfortunately they’re very simplistic: you can only say “rebuild all the indexes” or “reorganize all the indexes”. You cannot say, “If the index is 45% or more fragmented, rebuild it– otherwise do nothing.” If you don’t spend much time with SQL Server and you’ve got downtime available every weekend, this can be a decent option.
If you need to minimize downtime, custom index maintenance scripts are the way to go. Our favorite: Ola Hallengren’s maintenance scripts. These are super flexible, well documented, and … free! The scripts have all sorts of cool options like time boxing and statistics maintenance.
Some tips for using Ola Hallengren’s index maintenance scripts:
- Download and configure them on a test instance first. There’s a lot of options on parameters, and you’ll need to play with them.
- Get used the ‘cmdexec’ job step types. When you install the scripts you’ll see that the SQL Server Agent jobs run index maintenance using a call to sqlcmd.exe in an MSDOS style step. That’s by design!
- Use the examples on the website. If you scroll to the bottom of the index maintenance page you’ll find all sorts of examples showing how to get the procedure to do different useful things.
Find out when maintenance fails
Don’t forget to make sure that your maintenance jobs are successfully logging their progress. Set up Database Mail and operators so jobs let you know if they fail.
Tell your boss you did a good thing
Finally, write up a quick summary of what you did, why you chose custom scripts or maintenance plans, and why. Share it with your manager and explain that you’ve set up automated index maintenance as a proactive step.
Having your manager know you’re taking the time to follow best practices certainly won’t hurt– and one of these days, it just might help you out. (Even if it’s just by keeping everyone from following a red herring.)