T-SQL Tuesday #114 – The SQL Puzzle Party

T-SQL Tuesday logoThere were times when I tried to look for puzzles to solve, especially the T-SQL puzzles (what happened to the T-SQL Challenge site?). Now I don’t. Life is challenging as it is, especially if you work with SQL Server and really try to understand what’s going on.

So rather than coming up with some contrived problem for you to solve as part of this edition of T-SQL Tuesday (thank you Matthew McGiffen) I will share something that surprised me only last week. And yes, I have solved it already, and will be blogging more about it soon so no there is no big price for solving my production issue here 😉

Here is the scenario

There is a table that stores millions of records. It has a primary key, a date when a record was processed, a bit column indicating whether it was processed or not, and some text fields that are used for something, but in our example, it’s just data that takes space on pages.

There is also an application which is using nHibernate to generate a T-SQL query that retrieves one (just one at a time) records from that table where IsProcessed = 0. There are 10-50 records like that at peak times, in a table which holds tens of millions of records so making it very, very fast should be easy with a tiny little covering filtered index. Well… it turns out, SQL Server prefers to scan the clustered index instead.

Have a look

The challenge setup

use tempdb
go
drop table if exists dbo.LongProcessingTable
if not exists(select 1 from sys.tables where name = 'LongProcessingTable')
create table LongProcessingTable (
Id int not null identity primary key
,ProcessedOn datetime2 null
,IsProcessed bit null
,SomeData nvarchar(1024) not null
)

-- just some text to fill up the space on pages
declare @sometext nvarchar(1024) = (
select string_agg(convert(char(1),name), '')
from sys.all_objects
)

-- create just 100k records with some random date values
-- at this time all records are marked as processed
insert into dbo.LongProcessingTable(ProcessedOn, IsProcessed, SomeData)
select top(100000)
dateadd(second, -abs(checksum(a.object_id, b.object_id)%10000), getdate())
,1
,@sometext
from sys.all_objects a
cross join sys.all_objects b

-- now mark 10 rows as not processed
update d set IsProcessed = 0, ProcessedOn = null
from (
select top (10) *
from dbo.LongProcessingTable d
order by ProcessedOn desc
) d

Now the query:

declare @IsProcessed bit = 0

select top(1) Id, SomeData
from dbo.LongProcessingTable
where IsProcessed = @IsProcessed

The above query comes from the application and cannot be changed. It is what it is. And to help you start, here is the index I thought would work, but doesn’t.

create index IX_LongProcessingTable_NotProcessedYet
on dbo.LongProcessingTable(IsProcessed) include (SomeData)
where IsProcessed = 0

The index gets ignored and the server goes for the table scan instead.
Of course, there was somebody who discovered it earlier. I wasn’t all that surprised that Erik Darling blogged about it in 2015, 2017 and 2018 it turns out, he even says ‘IT IS KNOWN’… well, it wasn’t to me. But even know, with that knowledge, I still cannot change the query, so what can I do? How to make this query more efficient without changing it, and without creating a covering indexing on the whole table which can contain hundreds of GB of data just to get one row.

If you are still reading… well, enjoy the challenge. I will follow up with a few comments and a couple of my attempts at solving the problem later this month (hopefully).

 

A DBA’s thoughts on ORMs

What do you think about using NHibernate with SQL Server? What would you say if we were to use Entity Framework on our next project? What is your opinion on ORM frameworks?

I am a sort of DBA who spends a lot of his time working with developers. Deep in the implementation trenches, cutting code, trying to prevent any future ‘server issues’ by influencing the design at early stages of development. I find it much more efficient than waiting for them to chuck some code over the fence to us when there is very little we can do, but complain about them and get upset that somehow indexes and statistics don’t solve the problem. And so I hear those sort of questions a lot and hardly every I have the time to answer them in any other way than just to say ‘it depends’.

So here is an attempt on answering this question.

The developer in me wants to say:
Of course, use an ORM! Go code first if you can. It saves the time, it deals with the Object-Relational Impedance Mismatch problem, it keeps your code clean as there is no need for those strange looking data queries. All the code is in one place, one solution, easy to find, read and understand. It is data engine agnostic too, so we can deploy it on MySQL, Oracle, PostgreSQL or SQL Server. On anything really.

But then the DBA in me wants to shout:
Are you mad? Of course not! Don’t use ORMs. Ever. They produce unreadable, inefficient queries that are difficult to understand or optimise. The code first approach typically leads to inefficient schema. New database engine features are ignored because cross vendor compatibility is more important than performance. And don’t you see how those leaky abstractions of generic repositories you are using? Really, passing IQueryable to the business layer? Maybe you have the ability to run it on multiple data engine, but now your business layer depends on your ORM framework and the data model.. Read the Clean Architecture by Uncle Bob, especially the part about keeping frameworks at arm’s length.

And so the developer responds:
OK. So I will be more specific with my repositories… perhaps. Fine. But I’m not going to write any SQL statements. I don’t want any magic strings in my code with no support from the IDE. And no, no stored procedures. We cannot have logic split into multiple layers. All code needs to be in the repo, all code needs to be tested. Don’t you see, stored procedures just don’t fit in the modern software development cycle. Besides, we have developers who can write LINQ and don’t need to know any SQL.

But the DBA with a smug look on his face says:
Ha! That idea of abstracting away technology, so that you don’t have to understand it has been tried before. Sometimes it works, sometimes it doesn’t. What happened to WebForms. Wasn’t it the idea to hide HTML and JavaScrip to make web development easier for existing windows developers. How did that go?

And that’s how it starts again and again, and the discussion in my head goes on and on. But eventually I come to similar sort of conclusion time after time, and here is what I actually do. (It is a compromise on which both the developer and the DBA in me agree on, allowing me to stay sane).

  • For Proof of Concept work I use ORMs and the code first approach. That saves a lot of time and effort, and the code will be a throwaway anyway. My ORM of choice is Entity Framework but it doesn’t really matter.
  • I don’t spend much time thinking about data types. In most cases string defaulting to nvarchar(255) is good enough for a PoC.
  • I prefer to use EF Core as it supports in memory storage for even faster PoC development and testing.
  • Just in case it is not thrown away (as it should), I keep my architecture clean. I make sure to use specific repositories for data access, and that the repository abstraction is not leaking any implementation details. A repository takes and returns business objects and is using ORM framework internally only.
  • On projects which will not be thrown away I start with Dapper (a micro ORM) and stored procedures. It is a bit more work but forces me to design the data structures better, and offers a lot benefits for the future (more about it later in this post).
  • While I agree that logic should be in one place, there are different types of logic, and those should be implemented independently. There is UI Logic, there is Business Logic and there is Persistence Logic which I implement in a repository or in stored procedures. A good example would be a soft delete functionality.
  • All SQL code is kept in the same solution, is tested and deployed through the normal CI/CD channels using DbUp project.

So my answer is

Use ORMs as long as they work for you, but architect your code in such a way, that you don’t depend on them, and ready to ditch them when they start to cause more problems then they solve. Consider micro ORMs. Try Dapper.

Here are a few more benefits of using Dapper with stored procedures

  • Dapper has much smaller footprint than NHibernate or Entity Framwork.
  • Dapper is faster, almost as fast as a DataReader, when compared to full ORM frameworks. According to this at the moment of writing this post it is 10 times faster than NHibernate.
  • While being small and fast Dapper still takes the ORM problem away.
  • Stored procedures add some extra code that needs to be written, but allow access to the latest database engine features. In case on SQL Server those can be Hekaton (In-Memory OLTP), JSON or XML data types, graph structures, temporal tables, windowing functions and much more.
  • Stored procedures make performance troubleshooting and reviews much easier. For DBAs it is much easier to understand which part of an application creates the load and therefore what it is trying to do with well named stored procedure rather than a lot of auto generated SQL statements.
  • The cooperation with Database Developers is much easier, as they can easily identify queries that need to be optimised, and then improve them without worrying (too much) about any non SQL code.
  • Even if you don’t have DBAs and DBDs just now, you might in the future. If the business is successful it might be that you suddenly need to get somebody with those skills to help you. Having good structure, with a separated data layer will make their life easier, your bill lower and everybody happier.

Hashing values in SQL Server

TL;DR

  • SQL Server can hash values using some of the common hashing algorithms like MD or SHA.
  • It is possible to use XQuery in addition to XPath in XML value() function to do things T-SQL cannot do on its own.

The Details

Hash values or (hash codes) is what we typically use to store_passwords in databases. We use salt values too. Definitely, we don’t store clear text passwords. Right?

Typically during a login to an application, a password is combined with a salt value stored somewhere in a database and a hash value is calculated which is then compared with the hash value previously stored in a database. But there are other options. It is possible to calculate hash values directly in a database using T-SQL. It could be useful if a bulk update needs to be performed if you want to generate a lot of test users with predefined passwords during a database migration, but it is also possible to overwrite a password hash and gain access to the application as any user.

SQL Server (starting with 2008) has hashbytes function which can be used to calculate hashes using a number of different algorithms. The algorithms supported differ from version to version. The latest 2017 supports MD2, MD4, MD5, SHA, SHA1, SHA2_256, SHA2_512.

Let’s have a look at how to use it

declare @password nvarchar(32) = 'Secret1234'

select 'MD5' Algorithm, hashbytes('md5', @password) Hash
union all select 'SHA' ,hashbytes('sha', @password)
union all select 'SHA2_256' ,hashbytes('sha2_256', @password)

which produces something like this

Hashing results

You can see that the results are of varbinary type. That’s OK if your application is storing the hash in this format, but from my experiance, most developers will not know that a byte array can be stored in a database and they will convert it into a Base64 string.

SQL Server does not do Base64 encoding (not as far as I know) but it does support XML and XQuery, and they do encoding. So let’s use it.

declare @password nvarchar(32) = 'Secret1234'

select Algorithm
,convert(xml, N'').value('
xs:base64Binary(xs:hexBinary(sql:column("Hash")))', 'varchar(max)'
) Base64Hash
from (
select 'MD5' Algorithm, hashbytes('md5', @password) Hash
union all select 'SHA' ,hashbytes('sha', @password)
union all select 'SHA2_256' ,hashbytes('sha2_256', @password)
) t

Here I’m converting an empty string N'' to an XML type which creates an empty XML object which allows me to use it’s value() method to execute XQuery which is a functional language. In most examples of XML in SQL Server the only thing you will see is XPath in value() but XQuery can be used too.

The results are more what you’d expect:

Hashes to Base64

One thing to note. The hashing algorithms operator on bytes so are not only case sensitive but type sensitive too.

declare @password nvarchar(32) = 'Secret1234'

select
'varchar' Type
,hashbytes('md5', convert(varchar(32), @password)) Hash
union all select
'nvarchar'
,hashbytes('md5', convert(nvarchar(32), @password))

Have a look at the results. Only because type used is different, the hash is different too.

Hash values using different types

If you want to make it work with .Net make sure to use nvarchar as in .Net strings are Unicode.

How can it be useful?

  • It is possible to push hash calculation and comparison to database which means the correct hash and salt value are never loaded to the application memory.
  • It is possible to generate hashes for test users in database seeding scripts avoiding application doing row by row processing.
  • It is possible to generate hashes of objects to detect changes (although the checksum function can good enough and faster too).
  • It is possible to use XQuery to do things T-SQL cannot do.

But it brings some risk too
* It is possible to gain access to an account by updating values in a database even when using hashing with salt.

Connect to SQL Server on Linux

How to connect to SQL Server on Linux?
It really depends on what one means by ‘connect’.

Query SQL Server on Linux from Windows

To connect to the SQL Server instance running on Linux directly or in a container to run a query it is no different from connecting to one running on Windows. Just connect using the DNS name or an IP and a port number if it is not the default 1433. The only difference is that only SQL Authentication is supported so you will not be able to use Windows Authentication and your domain credentials.

Query SQL Server from Linux

Azure Data Studio (formerly known as SQL Server Operations Studio) can run on Linux offering the same user experience as on Windows. It is also possible to use the sqlcmd which works exactly the same as the one on windows. To install it on Ubuntu or any Linux flavour using apt-get package manager you can simply do

apt-get install mssql-tools

End then just use it the same way as you would normally do

sqlcmd -S localhost -U sa -P mysecretpassword

In the default SQL Server Docker image the mssql-tools are already installer but they are not added to the $PATH variable. You can either add it, or use the full path to execute it which changes the above command to

/opt/mssql-tools/bin/sqlcmd -S localhost -U sa -P mysecretpassword

Remote to Linux remote server

When managing a SQL Server instance it is sometimes necessary to connect to the operating system on which it runs. This is very different on Linux to running SQL Server on Windows. There is no remote desktop, and PowerShell doesn’t really work either. That’s where the typical Linux admin tools become necessary.

In a way the Secure Shell or SSH is the Linux equivalent of the Remote Desktop. There are many ways to do it, especially from another Linux box. Most of us, SQL DBAs, will be typically on a Windows machine. In that case a common approach is to use PuTTy. But if you use PowerShell and keep up with updates it is now possible to SSH directly for a PowerShell terminal too.

ssh mysqlonlinux.mydomain.com

or

ssh 10.11.12.13

Simple as that. You will connect with a specific user and now most of the admin commands will have to start with sudo which allows to execute them with elevated permissions. It’s similar to Windows’ Run as Administrator.

Connect to Docker container from the host

If you want to connect to a container running your SQL Server on Linux from the host use docker exec to start a bash shell.

docker exec -it sql1 bash

where sql1 is the name of the container hosting SQL Server.

You will connect with superuser permissions and sudo will not be necessary. The prompt will look something like this:

root@2ff21ad83800:/#

SSH into a container running SQL Server on Linux

If you cannot or don’t want to remote onto the host first to connect to the container running your SQL Server instance SSH is the way to go, again. By default, the SQL Server image does not have SSH server so you will have to install it or a custom image will have to be created. Once that is done it is no different from connecting with the ssh (or PuTTy) to any other Linux server. It doesn’t matter that it runs in a container.

Using SQL Server on Docker (1)

It is yesterday’s news. SQL Server runs on Linux, and on Docker too. There are plenty of blog posts showing how to install, start and connect to it too.

During last Data Relay, I have seen Mark Pryce-Maher‘s talk on SQL Server on Linux (twice). It was a good talk, but it followed the same pattern so not surprisingly a common question from the audience appeared to be ‘why would you want to do it?’.

I mean, the biggest claim to fame appears to be that SQL Server actually works on Linux. It doesn’t bring any new features. In fact, some features are missing. A good part of docker demos appears to be an exercise in futility too. One starts a container with SQL Server, creates a database, inserts some data then the container restarts and as if by magic the data is gone, the database never existed. So, indeed, why would you want to run an SQL Server on Docker?

This post is my attempt to address this question by showing how I use SQL Server on Docker and how I deal with the persistence problem.

First: Get it up and running!

If you want to know how to run and connect to SQL Server on Docker read the Microsoft’s Quickstart Document, or Andrew Pruski’s blog post about running vNext on Docker.

Making the long story short, assuming you are running Windows, have Docker installed, configured to run Linux containers and that you don’t have an SQL Server instance locally installed:

docker pull mcr.microsoft.com/mssql/server:latest
docker run -d -e ACCEPT_EULA=Y -e SA_PASSWORD=Secret1234 -p 1433:1433 --name sql mcr.microsoft.com/mssql/server:latest

That’s all you need to do to have the latest production SQL Server (Developer Edition) instance running on localhost on port 1433.

Creating a database that survives container restarts

If you have the container from the previous example running let’s stop and remove it first.

docker stop sql
docker rm sql

There is a number of ways to persist container data. For my purposes the simples appears to be docker volumes. You can read more about docker persistence on the official documentation site but for now let’s just create a local volume called sqlvolume

docker volume create sqlvolume

and mount that volume on the container

docker run -d -e ACCEPT_EULA=Y -e SA_PASSWORD=Secret1234 -p 1433:1433 --mount source=sqlvolume,target=/data --name sql mcr.microsoft.com/mssql/server:latest

Now, if you create a database with files on the mounted volume like so:

create database [TestDb]
on primary (name = N'TestDb', filename = N'/Data/TestDb.mdf')
log on (name = N'TestDb_log', filename = N'/Data/TestDb.ldf')

and you stop and start the container

docker stop sql
docker start sql

as if by magic, the TestDb database survived the restart.

Why I find it useful

So with some considerable effort, I was able to create a database that survives a restart of a container or that of a host system. Nothing a standard SQL Server instance running on windows couldn’t do! So why do I do it?

I write my blog posts on my laptop and I don’t want to commit to installing SQL Server on it. I want to be able to try out vNext while still having the ability to run demos against 2017 and I’m definitely not installing two instances on my laptop. Luckily I don’t have to. With the above setup all I have to do is run docker start sql and within seconds I have a local instance of SQL Server ready to serve my queries. When I’m done I do docker stop sql and it is as if the SQL Server was never there.