Azure Resource Locks are Your Friend in Development

One of the great advantages of the cloud computing is the ability to power off resources that are not in use to save some money. Sure, your production database servers should be running 24×7, but that VM, or SQL Data Warehouse you are developing against during the week? You can shut it down at 7 PM (1900 for the Europeans reading this) and not start it up. Azure even recently introduced an auto-shutdown feature for VMs.

Screen Shot 2017-03-24 at 8.55.37 AM

Unfortunately, there is no auto-startup feature, but that is easy enough to code using an Azure automation job.

This sounds great, can it walk my dog, too?

Unfortunately, there’s one problem with our awesome budget saving proposal. Sometimes developers have jobs that run beyond the time they leave the office. For example, last night at one of my clients a developer had an SSIS package running after he left, and it got killed when the SSIS machine auto-shutdown at 7. That isn’t good.

The solution for this is Azure resource locks—you can put a lock on any resource in Azure. A lock can do one of the two things—first there are delete locks which simply keep a resource from being deleted. It is not a bad idea to put a delete lock on all of your production resources to prevent any accidental deletion from happening. The second type of lock is a read-only lock, and these are a little more aggressive. You can’t do anything to a resource with a read-only lock—you can’t add a drive to a VM, you can’t resize, and most importantly, you can’t shutdown the resource.

You can use the portal, PowerShell, or CLI to create a lock. It’s a fairly simple construct that can be extremely beneficial. You can get current details for lock creation from the Azure Documentation.

My developers have access to the portal (thanks to role based access control and resource groups), so I’ve instructed them on how to place locks on resources, and how to remove them. As an administrator, you probably want to monitor for locks, to ensure that they aren’t left in place after they are needed.

You’re Speaking…and You Don’t Have Slides

I had this dream that other week. I was in the big room at PASS Summit, sitting in the audience. I was relaxed, as I thought I was presenting later in the day, when I quickly realized, due to the lack of speaker on the stage, that I was the next speaker, and the room was full. And I was playing with my laptop and I didn’t have a slide deck. In my dream, this talk was a 300 level session on troubleshooting SQL Server, something I feel like I could do pretty easily, you know with slides. Or a whiteboard.


I woke up, before I started speaking. So, I’m not sure how I would have handled it—interpretive dance? I’m a pretty bad dancer. One thing, I will mention, and I saw my friend Allan Hirt (b|t) have to do this last month in Boston—really good (and really well rehearsed) speakers, can do a very good talk without their slides. Slides can be a crutch—one of the common refrains in Speaker Idol judging is don’t read your slides. It is bad form—do I sometimes read my slides? Yeah, everyone does occasionally. But when you want to deliver a solid technical message, the best way to do that is telling stories.

I’m doing a talk next month in Belgium (April 10, in Gent), right before SQL Bits. It’s going to be about what not to do in DR. My slide deck is mostly going to be pictures, and I’m going to tell stories—stories from throughout my career, and some stores from friends. It’s going to be fun, names will be changed to protect the guilty.

So my question and guidance for you dear readers, is to think about what you would do if the projector failed and you did not have a whiteboard. I can think of a number of talks I can do without a whiteboard–in India last year, another instructor and I demonstrated Azure networking by using our bodies as props. What would you do in this situation?

Monitoring Availability Groups—New Tools from Solarwinds

As I mentioned in my post a couple of weeks ago, monitoring the plan cache on a readable secondary replica can be a challenge. My customer was seeing dramatically different performance, depending on whether a node was primary or secondary. As amazing as the Query Store in SQL Server 2016 is, it does not allow you to view statistics from the readable secondary. So that leaves you writing xQuery to mine the plan cache DMVs for the query information you are trying to identify.

My friends at Solarwinds (Lawyers: see disclaimer at bottom of post) introduced version 11.0 of Database Performance Analyzer (DPA, a product you may remember as Ignite) which has full support for Availability Group monitoring. As you can see in the screenshot below, DPA gives a nice overview of the status of your AG, and also lets you dig into the performance on each node.


There are a host of other features in their new releases, which you can check out some of their new hybrid features in their flagship product Orion. Amongst these features, a couple jumped out at me—there is now support for Amazon RDS and Azure SQL Database in DPA, and there is some really cool correlation data that will let your compare performance across your infrastructure. So, when you the DBA is arguing with the SAN, network, and VM teams about where the root cause of the performance problem, this tool can quickly isolate the root cause of the issue. With less fighting. These are great products, give them a look.

Disclaimer: I was not paid for this post, but I do paid work for SolarWinds on a regular basis.

SQL Clone—Win Fabulous Prizes!!!

Want a chance to win a really cool prize (5 Amazon Dots, and a copy of SQL Clone), while learning about a cool product from Redgate? My friends at Redgate are in the process of releasing a really awesome tool call SQL Clone. So what is SQL Clone?

Screen Shot 2017-03-06 at 10.29.34 AM

One of the challenges I’ve faced through my career as a DBA is the ability to have a reliable copy of a production database to test against. Sure, when I worked at Comcast we had dev, qa, test, and prod environments, but many of the smaller organizations I worked for and consult with, don’t have that luxury. Even when I was at Comcast, the overwhelming cost of storage really limited our ability to maintain consistent development platforms.

So what is the solution? SQL Clone offers a virtual copies of your  database, in seconds, whilst only taking up 40 MB of disk space. How does it do it? Redgate stores pointers to your database in the copy. This means you can have one copy of your database and multiple clones, so that different parts of your dev team can work together in harmony. This data will grow a bit as each developer queries the data, but if you dealing with very large (> 5TB) or even just pretty big databases, this can greatly reduce your storage cost and increate developer productivity.

To learn more about Red Gate SQL Clone, check out this video:

You can see Grant Fritchey yell at an Amazon Alexa to create a clone of a database. It’s that easy.

So for the contest:

I’d like you to comment on what skill you would build using the echo dot. This can be anything, but database related ideas are likely better, as Redgate will be the judge. Here are some of my ideas:

  • Build a database using SQL Clone Smile
  • Rebuild the statistics on the data warehouse
  • Yell at Joey to go ride his bike

    For legal purposes you can read the terms and conditions here.

Query Store and Availability Groups—Force Plan on Secondary Replicas

I’m still fighting with some challenges about inconsistent performance between a primary and secondary replica, so I’ve been waste deep in undocumented system views looking at temporary statistics. One of the things I thought about doing was talking advantage of the Force Plan option in the Query Store in SQL Server 2016.  If you are not familiar with this feature, it allows you to force a “preferred” execution plan. In this scenario, our query was running in about 20-30 seconds on the primary, and 20-30 minutes on the secondary. The plans were reasonably close, but I wanted to see what would happen if I forced a plan on the primary.

Primer about the Query Store and Availability Groups

Since readable secondary replicas are read-only, the query store on those secondary replicas are also read-only. This means runtime statistics for queries executed on those replicas are not recorded into the query store. All the stats there are from the primary replica. However, I wasn’t sure what would happen if I forced a plan on the primary—would the secondary replica honor that plan?

Let’s Find Out

The first thing I did was to query the query store catalog views to verify that the plan was forced.


I have to copies of the forced plan. If I run an estimated query plan on the primary, I see that plan is forced. You can see this by looked for UsePlan in the XML of the plan.


I did the same thing on the secondary (in the case of the secondary, we are looking at the actual plan, but it doesn’t matter).


You will note that there is no UsePlan. There are extended events and a catalog view that reflect plan forcing failure (Grant Fritchey wrote about this behavior here), While, I wouldn’t expect the catalog view to get updated, I was hoping that the Extended Event might fire. It did not.


The query store, as awesome as it is, doesn’t really do much for you on readable secondary replica. It does not force plans, nor does it record any of your data.

Thanks to Grant Fritchey and Erin Stellato for helping with this post!

The Curious Case of the HTDELETE Wait Type

I was working with a client this week and we encountered very long wait types on “insert as select” queries that were part of their data delivery process. This wait type isn’t documented very well, SQL Skills has it documented here  and mentions this:

“Typically these waits occur when queries involve columnstore indexes, but they can also occur without columnstore indexes being involved if a hash operator runs in batch mode.”

Nacho from Microsoft also has a blog post on it here. My theory was that a bad hash join was taking place and causing the wait.

Isolating the Plan

The thing that was very curious about the situation is that waits were only occurring on the readable secondary replica. At first, I tried to examine the query store to try to understand if there were multiple execution plans for a given query. The one problem with that is the readable secondary copy of the data is read-only, which means on that secondary replica you only see the query store data from the primary replica. If there was a plan that was specific to the secondary, I’d have gather them from the plan cache on the secondary. (Thanks to Erin Stellato (b|t) for this idea). There was one other problem—the code in question was executing as dynamic SQL from a stored procedure when meant it was always getting a new execution plan.

Ultimately I think any thought of the readable secondary having a vastly different plan was a red herrings. Statistics are going to be the same on both instances, and if there were a missing statistic on the secondary, SQL Server would create it in TempDB. Anyway, columnstore indexes don’t use statistics in the traditional sense.

Fortunately I was able to catch a query in the process of waiting on HTDELETE, so I no longer had to look for the needle in the haystack, and I could get to tuning the plans. I was able to grab the SELECT part of the query and generate an estimated plan on both the primary and secondary nodes. The plans were virtually the same on both nodes, with just a minor difference in memory grant between them.



The query is here.

SELECT    Object4.Column1
FROM    Database1.Schema1.Object5 Object4
    INNER JOIN Database2.Schema1.Object6(?, ?) Object7 ON Object4.Column2 = Object7.Column3
    INNER JOIN Database2.Schema1.Object8(?) Object9 ON Object4.Column4 = Object9.Column4 
    INNER JOIN Database1.Schema1.Object10 Object11 ON Object4.Column5 = Object11.Column6
    INNER JOIN Database2.Schema1.Object12(?) Object13 ON Object11.Column7 = Object13.Column7
WHERE    1 = 1
    AND Object4.Column8 >=  ‘01-Jan-2017’

The pattern here was that we were taking all of rows of an ID field in a columnstore index with about 350MM rows and joining them to a function that has 3500 rows. My gut instinct was this was a bad match for batch mode hashing. Additionally, SQL Server was recommending I create a b-tree index on the large columnstore table. there was a key lookup in the plan that I wanted to eliminate, but my hunch was that this join was causing the waits.



The Solution

So before I created the index, the query was taking at least 2-4 minutes, when it wasn’t getting hung on the HTDELETE wait. After I created the first index, we got done to about 15 seconds. SQL Server then recommended that I create another index on one of the join tables, which brought my query time down to sub-second. The plan looked a lot more traditional and had lots of my favorite operator INDEX SEEK.



The Moral of the Story

Sometimes you need non-clustered indexes on columnstore indexes. It stinks, because they do add space, but its hard to argue with a performance gain like this. I need to email some friends on the product team to ask, but I’m therorizing that the join was super expensive and causing the query to hang. Anyway, the real answer is to never stop tuning and trust your instincts.

Thanks to Sentry One for making Plan Explorer Free. I used it for the screen shots and anonymization in this post.

SQL Bits Precon: SQL Server on Linux—A Brave New World

My first European trip this year includes two of my favorite bike races, the Tour of Flanders (or as my Belgian friends call it, the Ronde van Vlaanderen) and Paris-Roubaix. In the week between on Wednesday, I will be doing a full day of training at SQLBITS in Telford. I’ve had the good luck to be working with SQL Server on Linux since very early days of the development process, and am looking forward to sharing that knowledge with attendees.



In this full day training session you will learn about the Linux operating system. Some of the topics we’ll talk about include:

  • Operating system architecture
  • Security Model
  • File manipulation
  • Common commands and using them together
  • Bash and Korn shells
  • Shell scripting
  • File systems and Volume Managers
  • Clustering in Linux
    You will also lean about technologies in SQL Server.
  • Monitoring O/S performance 
  • Proper SQL Server configuration
  • Automating deployment
  • Deploying High Availability and Disaster Recovery
    Whether you are new to Linux and UNIX and want to get started, or if you can awk and sed your way through a maze, this precon will have something for you.

DBCC Clonedatabase and Very Large Databases

One of the recent feature introductions to SQL Server is dbcc clonedatabase, a feature that lets you create a “data-less” clone of you database. All of the statistics and objects come into your cloned database, however none of the data does. This is perfect for development or performance tuning exercises, where you want all the metadata, but do not want the security risk of dealing with production data.

Recently I had the opportunity to use clonedatabase on a very large database. I was concerned about the size of the data files and how this would impact space on my volumes. Books Online is fairly clear, but I wanted to see for myself.

Note All files in the target database will inherit the size and growth settings from the model database. File name convention: The file names for the destination database will follow the source_file_name _underscore_random number convention. If the generated file name already exists in the destination folder, DBCC CLONEDATABASE will fail

So my thought in reading that, is that the same number of data files will be created in the clone, just with the settings in model. Let’s test that out.

The first thing I did was create a new database, and then add a few data files to it. I made them 20 MB, which is a different size than model—just for testing purposes.


Next, I ran the clone database command.


Then connect to the clone and look at the data files


I can see that all of the files were created, in the same location as the files on the source database, except with the size settings of model. While this shouldn’t be a big deal for most, if you do like I recommend and make model a reasonable size for your environment, and you happen to be tight on drive space, you could fill up a volume. So just be aware when using clonedatabase particularly with databases that have a lot of data files in them.

An “Ask” for Microsoft—A Global Price List

And yes, I just used ask as a noun (I feel dirty), I wouldn’t do that in any other context, but this one. In reviewing my end of year blog metrics, my number one post from last year was a post that listed the list price of SQL Server. I wrote this post because a) I wanted clicks and b) I knew what a pain it was to find the pricing in Microsoft documents. However, the bigger issue is that to really figure out what a SQL Server cost, you need to go to another site to get Windows pricing, and probably another site to find out what adding System Center to your server might cost.

This post came up because Denny and I were talking the other night, as someone had posted to the Data Platform MVP list asking how much the standalone R Server product cost. We found a table on some Microsoft site:


I’m not sure what math is required to translate “Commercial Software” into a numeric value, but it is definitely a type conversion and those perform terribly. Eventually I found this on an Azure page:

This image is charged exactly like SQL Server 2016 Enterprise image, but it contains no Database elements and has the core ScaleR and DeployR functionality optimized for Windows environments. For production workloads we recommend that you use a virtual machine size of DS4 or higher.

This leads me to believe that R Server has the same pricing as SQL Server, but with the documents I have I am not certain of that fact.

What Do I Want?

What I want, is, a one-stop shop where I can find pricing for all things Microsoft, whether they be Azure, On-Premises, or Software as a Service. At worse it should be one click from the product name to it’s pricing page. Ideally, I’d like it all in a single table, but let’s face it, software pricing can be complex and each product probably needs it’s own page with pricing details.

The other thing that would be really cool, and this is more of an Azure thing, is to have pricing data built-in to the API for deploying solutions. That way I can build pricing based intelligence into my automation code, to rollout cost optimized solutions for Azure.

Anyone else have feature suggestions?

Updated: Jason Hall has a great comment below that I totally forgot about. Oracle has a very good price list (it definitely wins the number of commas award) that is very easy to access. So dear readers in Redmond: Oracle does it, we you should too!

Updated: There is some of this available in Azure. It’s not perfect though. Amazon just announced enhancements to their version of this service.

SQL Server on Linux Clustering—A Few Other Notes

So I was chatting with fellow MVP Allan Hirt (b|t) about the cluster build that I wrote about yesterday, and I had a few more realizations about the Linux HA process as it stands right now. I haven’t talked to the the Linux product team at Microsoft about this, but I hope to in the near future to get a better idea of where things are headed. So these are my notes as of now, strictly relating to failover cluster instances (FCI), AlwaysOn Availablity Groups are coming, but are not in the latest CTPs of SQL Server on Linux.

It was faster than building a Windows cluster

It took me a while, I laughed, I cried, I cursed a lot, but if I look at the time it took for me to actually build the cluster and install SQL Server, it was a much faster process. Much of this comes down to the efficiency of the SQL Server installation process on Linux, which is as simple as running yum install mssql-server (mostly). Which leads me to my next point..

Installation options would be nice

The cluster building process is a little kludgy. Basically, you install two standalone instances of SQL Server, and then remove the data files from one them, and copy them into your NFS share. Having the option to do the equivalent of an “Add Node” install, would mean you wouldn’t need to worry about cleaning up your second node.

There’s no cluster validation, explicitly

This is a bit scarier, or easier depending on your view point. There are tests at various parts of the process to make sure things are working. For example, the first step of building your Linux cluster is to authorize the nodes to take part in the cluster, which validates certain security and network settings. However, the storage validation consists of starting and stopping SQL Server on each node to make sure it can talk to the storage and startup. Given that Microsoft doesn’t own the clusterware for this solution, I’m not sure how much they can enhance that, or if they will. This is a good open question.

There’s no dns

(Happy Late Birthday Kris!) One interesting thing I realized after talking to Allan was that I did all of my networking setup through the /etc/hosts file on each individual node. I remember doing this for RAC, and I think it may be a requirement of Pacemaker, but you will still want to make a DNS entry for your cluster identifier. When you do this on Windows, if you are using Active Directory for DNS, the installation does this for you. Not in Linux, you will need to do this yourself.

Screen Shot 2017-01-04 at 11.40.26 AM

Get comfortable with command line and scripting

There’s no cluster wizard to get you through the process. I think this isn’t a huge deal—Denny and I were talking yesterday about how relatively easy it would be to script the whole process in bash (I’m holding off until I find out if Microsoft is doing this), and most Linux sysadmins are really comfortable with writing bash scripts. But if you aren’t comfortable with Linux and the command line, now is the time to brush up, before things go prod.


We are in the very early days of this process, there is much that will likely change. From a functional and conceptual perspective, this is very similar to the way a SQL Server Failover Cluster works in Windows, but the implementation is quite different. I’d like to see things resemble Windows a bit more, at least from a SQL Server perspective, but we’ll see where the product heads.

%d bloggers like this: