The below is an extract from my presentation at SQL Saturday 252 on high availability and disaster recovery.  In this section we talk about the basic process of backing up databases and whether this is enough to meet your requirements.

Let’s leave aside all the cool options you get natively and talk about something more mundane.  Backups.  Regardless of what else you have in place, you still want to be doing your backups regularly.  If everything goes completely pear shaped, if all your best laid plans fall apart, so long as you have a native SQL backup then you still have your data.  Without all the other options we are going to be discussing you can still recover from a serious outage if you have a secured backup.  There might be some trial and error if you haven’t documented users and settings, it might take some time and whenever you resort to restoring from backup you are making a call on acceptable data loss – but you haven’t lost everything.  You can recover.

To make it easier on yourself there’s other things that are critical to the application using the database you should be documenting and regularly scripting as well.  What version and patch level is the server? Are the logins being scripted?  What about agent jobs?  Linked servers?  SQL config settings?

I’ve come across businesses whose only recovery options was from their scheduled backups and they were not interested in any additional options.  Sometimes it’s a little surprising considering the nature of their businesses.  Some are financial and any dataloss would be measured in actual dollars, actual transactions being unrecorded and sold items not being sent to customers who would probably not be too happy about that.  One in particular springs to mind where the company had a defined and documented RTO and RPO.  It was backed by SLA’s around those RTO’s and RPO’s and they could take a day of downtime and lose up to 24 hours of data and still meet those targets.  In a disaster they had a documented build process and they would simply install(from scratch) a new server, install and patch SQL, restore the databases, reapply their standard permissions and repoint their application server.

This all raises the question of whether just taking your backups qualifies as Higher Availability or Disaster Recovery.  To me…..so long as you are able to meet the RTO and RPO then it absolutely is.  It’s also going to be largely dependent on the data sizes you are talking about.  When you are talking megabytes or low gigabytes you can probably get away with this sort of strategy.  But when you are talking terrabytes just the time to retrieve and copy your backup to a new server will probably blow your RTO, let alone restoring it and hoping like hell it works right on the first attempt.  It comes back to the process of establishing those RTO and RPO values.  If everyone agrees on values that are very high, and they will actually admit they agreed to them if there really is a disaster, then using a standard backup strategy as your only form of DR\HA is fine.  It’s something that should be done even if you do have yourself covered with the other HA\DR options because it gives you a backstop should everything else go wrong.

Now having said that…….and this applies to everything I’ll be saying about the other technologies as well…..test it.  When you start putting together your backup strategy you need to think backwards.  Think of the process you are going to go through to restore everything.  Not just what do you need, but what order are you going to bring it back.  And document all the steps, because it may ultimately not be you doing the process, and it should be a process that more than one person can pick up and run with.

One of the other speakers at SQL Saturday was Argenis Fernandez.  I think the first blog post of his I ever read he was talking about a disaster situation where the DBA was confident he had everything in space.  Regular backups – taken off to tape each night – tape stored securely in a safe offsite.  Unfortunately he never tested the restore process.  So when a disaster rolled around and he put the tape back in to grab the backup from – the tape was completely blank.  The magnetic lock on the safe erased it.  You get a lot of peace of mind from knowing that the process that you are going to use to recover your system has been tested and works.

(EDIT – For the life of me I can’t find the post from Argenis I refer to above.  So perhaps I dreamt it).

(EDIT 2 – Here I am 3 years later re-writing my presentation from 2013 and I found the post I incorrectly attributed to Argenis.  It was actually written by Anthony Bartolo and can be found here.)

The below is an extract from my presentation at SQL Saturday 252 on high availability and disaster recovery.  This post deals with the concepts of RTO and RPO.

Let’s start by covering off just what we are talking about when we are using these terms – Higher Availability and Disaster Recovery.  Basically what we are referring to is a couple of acronyms – RTO and RPO.  That’s Recovery Time Objective and Recovery Point Objective.  Or put another way, how long can a service be down, and how much data loss is acceptable following an outage.

Now, in a lot of places these aren’t defined.  They are replaced by things like “best efforts” and the expectations are that you’ll get things up as soon as you can and lose as little as you can.  This is a dangerous mentality, because it should be the expectation anyway.  If you have a recovery time objective of an hour, and you get things fixed after 15 minutes, you aren’t going to look at your watch, see you’ve got another 45 minutes before your RTO runs out so go get a cup of coffee before you throw the switch.  You’re gonna turn stuff back on and then get the coffee.  Ditto for the RPO.  If you can retrieve all the data, you will do that, not delete or not restore up to the defined RPO.     Of course you are going to do things as quick as you can and lose as little as you can.

When I say it’s dangerous I’m looking at the outage that drags out.  If things take an hour to bring back online, or if 15 minutes of transactions are lost….or if you don’t know how much data is lost…then you put yourself in a position where someone can come back and question if your “Best Effort” was good enough, or perhaps if someone else might have done things better.

So you need to get these things defined by the business, or the application owner etc.  If they can’t give you hard numbers, or if they just drag the chain on doing it, then you need to give the numbers to them.  Tell them…”Okay, right now if this system goes down we stand to lose 15 minutes data and it’s take me an hour to bring everything back up.”  If they accept that – fantastic, you now have a documented RTO and RPO.  If they don’t then you’ve started the conversation about what additional resource you need to hit a different target.

Another important consideration for the recovery time objective is when the clock starts.  If you have an RTO of 1 hour – is that one hour from the time the issue arises, one hour from the time an automated alert is sent.  1 hour from the time you see that alert, or one hour from the time you actually get onto the server to look at the problem or one hour from the time the decision is made that you need to invoke DR.  Obviously this matters.

The bottom line is that if you don’t have those two numbers defined then you cannot define whether your higher availability and disaster recovery strategy is a success or not.

A note on ‘uptime’

Now when we define successful uptime it’s often defined as a percentage or number of 9’s.  Let’s take a quick look at that.  Let’s say you go to your boss and say we had 95% uptime.  That’s a nice number right?  On a test you’d be happy with a 95?  But in uptime terms that translates to 72 minutes downtime every day, or over 18 days a year of downtime.  So the more 9’s we add the narrower those allowances become.  At 99.999% downtime we are talking about less than a second per day and just over 5 minutes per year.  So basically you are talking one unscheduled restart a year.

What Disaster?

The final thing I want to talk about before we move on is the different types of disaster.  You need to think of HA and DR basically like insurance policies.  I live in Christchurch, which got rather smashed up in a recent earthquake, and once upon a time insurance policies were simple here.  You insured your house for a lump sum, and so long as you kept up your regular maintenance(payments) then you knew that should disaster strike your insurance company would come along and you’d have your new house up and running within a sensible time(Currently insurance companies are defining ‘reasonable time as around 5-10 years).  This is an example of responding to setting an RTO RPO with “Just do your best”.  Who’s to say 5 years isn’t reasonable if it’s not defined?  Nowadays you have to provide all types of addition information on that insurance policy – geotech reports and the like, because the Insurance companies have discovered that to replace a $400k house can often cost more that $400k.

In the SQL world it’s the same.  We need to consider more than the SQL server when we are talking about these things, because SQL Servers are not independent of everything else in your infrastructure and they are affected by different types of ‘disaster’.  You can have your server go down.  You can have a patch fail.  You can have your service account password changed and prevent SQL from starting.  You can have network or active directory problems.  Or you can have a data issue, where something that shouldn’t have been deleted is, where a “WHERE” clause is missed when executing a query.

And don’t forget this(We are in Christchurch so we are thinking of big natural disasters, and we are in IT so we know that sometimes hardware and software don’t do what you expect):  But not all disasters are accidents.  There are people out there who can and will deliberately do damage to your system – and anyone who came along to the SQL User Group a few months back will be able to see just how easily that can potentially happen.

So as we move through the rest of this presentation remember…..there are different sorts of disaster, and we need to consider those in our HA and DR strategy.

Having a record of the database backup history is particularly useful.  I use the following script a lot, but I find it’s main use is when I am logshipping and SOMEBODY breaks the recovery chain by taking an extra log backup.  This can give you an idea of who did it and where the log backup went.

SELECT TOP 50 bs.database_name, m.physical_device_name,
bs.user_name, bs.backup_start_date,
CASE bs.[type]
WHEN 'D' THEN 'Full'
WHEN 'I' THEN 'Differential'
WHEN 'L' THEN 'Transaction Log'
END
AS BackupType
FROM msdb.dbo.backupset bs
INNER JOIN msdb.dbo.backupmediafamily m ON bs.media_set_id = m.media_set_id
WHERE bs.database_name = 'AdventureWorks'
ORDER BY backup_start_date DESC, backup_finish_date
GO

If you are just after the last backup date of each database take a look at this one:

select T1.name as [DatabaseName],
MAX(T2.backup_finish_date) as [Last Backup Date]
from sys.databases T1 LEFT OUTER JOIN msdb.dbo.backupset T2
on t2.database_name=t1.name
where t1.name not in ('model','tempdb')
and t2.type in ('D', 'I', 'G','F') — Data, incremental, file/filegroup or file/filegroup incremental
group by T1.name
order by t1.name

The script I use to monitor fragmentation is a very slightly modified version of the one found here which gives me a little additional information and excludes a bunch of indexes I probably don’t care about:

SELECT OBJECT_NAME(ind.OBJECT_ID) AS TableName,
ind.name AS IndexName,indexstats.index_type_desc AS IndexType,indexstats.page_count AS [Page Count],
indexstats.avg_fragmentation_in_percent
FROM sys.dm_db_index_physical_stats(DB_ID(),NULL,NULL,NULL,NULL) indexstats
INNER JOIN sys.indexes ind
ON ind.object_id=indexstats.object_id
AND ind.index_id=indexstats.index_id
WHERE indexstats.avg_fragmentation_in_percent> 10
and indexstats.page_count>500
ORDER BY indexstats.avg_fragmentation_in_percent DESC

According to Paul Randall’s blog post on Where the Books Online Index Fragmentation Thresholds Come From, the generally accepted guidelines are to rebuild indexes larger than 1000 pages that have more than 30% fragmentation.  So in the query above I have trimmed out indexes with less than 500 pages and included page count in the output.  You can go to 1000 pages if you want, but I like to start with a more aerial view when I have to troubleshoot index fragmentation.  It’s also worth reorganising Indexes which are fragmented below 30% – according to the above article anywhere between 5 and 30% should be reorganised.

Paul also makes a really important point in the article quoted above, and that is that these numbers are made up to give us a starting point.  So use the script above freely, but don’t forget there are a lot of other factors involved in whether the indexes it returns are causing you problems, or your constant attempts to get them sitting nicely are actually doing nothing to improve your server performance and are just sucking up your time and effort.