I was having this conversation with coworker last week, and it seemed like the 100th time I’ve said this.  Then it popped up on StackOverflow this morning, and that got me typing, so here’s yet another salty developer’s rant on the topic.

There is no question that NULLs should be avoided when possible, because of the problems they can introduce several problems with indexing and query syntax. 

Due to the problems that NULLs can introduce, there has been a push to stop using them.  However, like most such movements, it has gotten out of control, to the point that some people fanatically insist that NULLs should never be used in the database.  I was in this camp for a lot of years, before I found it to be a little overzealous.

The answer is somewhere in between.  NULLs should be avoided whenever possible, but there are indeed valid business reasons for storing them. 

Often times you need to store an optional numeric value, and many people will tell you to just store a zero to indicate "no value", but this is an even worse antipattern of storing a magical value that really means something else.  What if then you can’t use zero for one field, because zero is considered a meaningful value as well, or because this value is being used as the multiple of a divisor, so you need to use some other magical value (-1?) for that field in just this case, and now you have a worse problem, because all of your optional numeric fields are now behaving differently.  Oye.

Dates are an even more compelling candidate for nullable fields.  The standard "no value" that people use in .NET is the default unassigned DateTime value, which is DateTime.MinValue, or more specifically January 1, 0001.  However, you cannot write this magic value into SQL database because the default minimum value for a SQL Server DATETIME field is January 1, 1973.  You now have to have something that checks that you are translating those values properly as they are written to and read from the database, and you have to have defensive coding all over the place that checks for whether your date fields are less than SqlDateTime.MinValue, instead of just checking whether they are equal to DateTime.MinValue.  Double Oye.

My preference is to deal with the values as they truly are, and not to build a lot of artificial constructs to hide the true meaning and usage of the field.  If a field may very well not have a value, make it nullable, and make it nullable in your application objects as well (if you language supports such a thing).  Then, anytime you are using that field, you are required to consider what should be done in the case where it is NULL, but that is actually a good thing.  Generally I am opposed to making developers waste brain cycles on unnecessary code complexity, but that is because it steals focus away from the true business problem being solved; however, in this case, the lack of a value IS part of the true business problem and must be thought through.  If you are just defaulting these values, then the developer writing a formula or algorithm will be less likely to think through those edge conditions where the values are missing, and may not even realize at the time that it is a possibility that those values are missing.

 

At the very least, it’s probably not nearly as good as it should be.  As a pathological job-hopper who sees one company after another wasting hidden hours and days on insufficient database change control strategies, I’m sick of it.

So how does my company handle database changes?  I’ve asked this of potential employers many times, and I usually just get blank stares, or a vague answer along the line of “uh… with scripts?”  Sometimes, if they are honest, they just reply “poorly.”  There’s not even an explicit Joel Test for it.  Considering how bad the situation is at most companies, the real test is not whether their database change management is any good, but just whether they are willing to recognize how problematic it really is.

Given how much thought and effort goes into source code control and change management at many of these same companies, it is confusing and a little unsettling that so much less progress has been made on the database change management front.  Many developers can give you a 15 minute explanation of their source code strategy, why they are doing certain things and referencing books and blog posts to support their approach, but when it comes to database changes it is usually just an ad-hoc system that has evolved over time and everyone is a little bit ashamed of it. 

Hopefully we can get to the bottom of why this is and come up with a few ways to make it better.  It’s not like the database is important, is it?

 
I like you, but not in that way

So why don’t developers think about databases the same way that they think about source code? 

The most obvious reason is that database changes are fundamentally different in many ways.  While source code is usually just a collection of files that recompiled, versioned, and released at any given time, databases are much more temperamental.  They have existing data, and they have history.  Sure, source code has history, which you can review for reference purposes, but in databases the historical lineage of a table is actually very important.  In C#, if you added a field to a class, and then a week later someone else changed the type, and then a month later someone else changed the name, it usually doesn’t really matter too much when or in what order those changes happened, all that matters is the current state of the code and that everything builds together and works in its latest state.  However, if you were to do the same thing to a new field in a table, it definitely makes a difference, because there are data implications at every step of the way.  This alone scares a lot of developers away from maintaining the database.

To many developers, there is something fundamentally uncontrollable about databases.  They don’t fit into the safe model that we’re use to.  Managing those changes is definitely introduces new challenge, and many developers just don’t want to be bothered.

 

Doing Business As

Another major reason for the difference is just cultural.  Developers want the database to work well, and they may even like tinkering around with some stored procedures from time to time, but at the end of the day they like to be able to wash their hands of anything that slightly resembles being a “DBA”.

When you compare the DBAs (database administrators) and SCMs (stupid code monkeys), often times there is just a different way of looking at the world.

Developers often see themselves as the general of a utilitarian army of code that they build and train and then order off to carry out their mission, and they usually won’t hesitate to trash it all and replace it if it proves itself unable to complete the mission.  DBAs on the other hand are used to dealing with gentle lion that could easily lose it temper and kill everyone in the room if it’s not treated with kindness and respect.  Developers often have the option to wipe the slate clean and start over, and usually want to, especially when they are dealing with someone else’s code.  DBAs however are stuck with the original version and they need to keep it alive and make it work until the point that the data can be migrated to a better system, and we all secretly know that is never going to happen.

 

Code source control is an over-ripe banana

As developers, we are awash with tools and techniques for managing our source code changes.  We have a seemingly endless number of source control platforms to choose from, each with more features than we could every possibly need or even understand, and some people build entire careers around defining complicated branching and merging strategies for massive codebases.  I figured for a while that the ideas of practical source control where pretty well nailed down, but then you read Eric’s Sink’s posts on Distributed Version Control  Systems, and get an idea of how quickly the field is continuing to evolve.

At some internal IT department, a group of developers is toiling away on a set of dull yet fabulously enterprise-y software that five people will use.  Jimmy checks his code changes into his company’s version control system of choice, where it is automatically held at the gates until it is code-reviewed by a senior developer, and then it is checked into the team’s current release branch.  Meanwhile the continuous integration build server will download, compile, package, and unit test the code to make sure that Jimmy hasn’t broken anything, and that nobody else’s changes broke Jimmy.  The package code is then migrated through an integration environment, QA environment, UAT environment, and staging environment on its way to production.  All the while, as the changes are validated from one environment to another, version numbers are automatically assigned to allow anyone to trace back the exact revision for a given build, and the corresponding code changes slowly work their way towards the code-release promised land, the Main Trunk.  Those branches can get pretty damned complicated, even when everything is going smoothly, and it never goes smoothly.

Code branchs can get  

Hopefully this is a process that Jimmy’s company evolved out of mixture of necessity and forethought over the years.  The other, less attractive, and far more common scenario is that the company hired an astronaut trapped in a developer’s body, bored with his work and not bound by any sense of urgency, who assured the company that he was going to implement a lot of big complicated processes because that’s just want professional companies do. 

In the end, a whole lot of people are paying a whole lot of attention to managing the source code.  Hopefully at your company, you are paying attention to this to.

 

Databases, the redheaded stepchild of source control

Now ask yourself, how does Jimmy handle database changes? 

For example, say that all database changes need to be checked into an isolated directory in source control; after all we’re not savages.  However, since they don’t really “build” a database, that a directory which is ignored by their continuous integration server.  This in turn breaks the unit tests that are pointing to the integration database server, so Jimmy then needs to run those scripts manually in the integration environment. 

In this process, Jimmy sees other database scripts that were checked in recently in the same source control directory, but he has no way to know which scripts have already been applied to the integration server.  For the briefest moment, Jimmy considers applying those scripts as well, just to make sure that the integration server is fully up-to-date, but then he realizes that he can’t be sure which scripts have already been run without manually comparing the schema and scripts to see which have been applied, and this would make Jimmy the defacto owner for any issue that arise because of it.  With his own tasks and deadlines to worry about, Jimmy doesn’t have the time or patience for this silliness, so he just deploys his scripts, forgets about the others, and hopes for the best. 

It’s worth noting here that this is the kind of things that can silently kill software quality.  A motivated developer just tried to make things a little better, but the process was so frustratingly inadequate that it was impractical for him to do so.  Software companies depend on their developers taking the initiative to improve things, and when they are discouraged from doing so, either by person or by process, the company will slowly slide backwards into mediocrity, and it will drag every developer’s morale with them.

 

Now once Jimmy makes the changes to the integration server database, that also breaks some other developers that have been using that server for their development.  Those developers now need to stop and download the latest code changes to get back in sync, cursing Jimmy’s name the whole way.

Anyhow, during the next deployment to QA, someone needs to remember that these specific changes need to get deployed.  Since there is no defined strategy for tying database changes to code changes, every time code is deployed there is a little bit of confusion around exactly which database changes need to be released, which were already released, and what order the scripts need to be run in.  Jimmy is getting upset.

Another darker possibility is that instead Jimmy needs to submit his changes to the database review board, a collection of detached idealists, college professors without the college, who will criticize every aspect of the scripts in order to justify their existence, but will not really offer any true value because they don’t understand the business problem that needs to be solved, nor do they appreciate the application considerations beyond the database that need to be satisfied.

One of the long term impacts of this is that Jimmy will look for any possible way to accomplish what he is trying to do without making database changes, because, in his colorful youthful vernacular, “making changes to the database is a sodding bitch.”  And if he does indeed need to change the database, he’ll try to accomplish it just by changing stored procedures, because changing table schemas are even worse.  In the end, he’s definitely not trying to find the appropriate solution to the problem; instead he backed into a situation of being forced to find a “good-enough” solution that will minimize his hassle, regardless of the downstream impacts. 

From now on, he’ll look for any way he can accomplish it by only changing stored procedures and not changing the underlying schema.  If he’s lucky (sort of), he’ll find a way that he can just kludge the stored procedures to work around the problem for now, and let it be someone else’s problem to solve later.  He has long since given up trying to find the “right” solution, because it is so exhausting the current state of things is so busted up that it’s not even worth trying anymore.

Further complicating the issue, some developers and DBAs make one-off changes in the QA and production environments without going through source control.  Either they need to make an emergency fix in one of the environments and forget to go back and add the scripts to source control, or worse they just don’t believe that databases should be under source control (I’ve seen this attitude far too often from some DBAs, because they can’t stand the idea of being forced to use a source control system managed by some other developers, just so that they can make changes to their own database).  Pretty soon, every environment is a perverted branch of the one true database, and trying to identify why a test works in one environment and fails in another quickly becomes a nightmare.

 

Some day, things will be different!

So what’s the solution?  Well, in my humble opinion, we need something like this:

  • The database must be versioned, so that it is easy to tell which changes have applied and when they were applied.
  • All database changes must be checked into source control.
  • All database changes can tied to the code changes that they affect, ideally checked into source control as a part of the same changeset transaction.
  • The database changes are built along with the code changes.
  • The database changes are deployed along with the code changes.
  • The continuous integration server must be able to build and update its own copy of the database, so that it can run automated tests of code and scripts that are checked in at the same time.

I’ve seen some companies that have had home-grown utilities that come close to accomplishing this, but in the end they all fell just a little bit short, which is not too bad.  However, the vast majority of companies I’ve seen were not even in the ball park.

Some of you are probably asking, “doesn’t Visual Studio Team System” do this?  Yeah, I think so.  Probably, but who knows.  Honestly I tried working with it a few times, and it caused me nothing but problems.  Sure, I could spend a lot of time mastering all the quirks, but I’m looking for something a little bit more accessible here.  The underlying concepts are hard enough; we need an approach that simplifies it, and I just don’t think that VSTS accomplishes that.  More importantly, and also along the lines of accessibility, VSTS costs a fortune, and so most developers will never have access to it, so I’d like to fine something that can work for the other 95% of developers out there that are stuck using reasonably-priced tools.

What about Red Gate’s SQL Compare and SQL Data Compare products?  Absolutely, they tools are indeed awesome products that can prove invaluable in many cases.  However, they are often leveraged once an unknown number of database changes have already been made by an unknown number of people.  However, I think the better solution, whenever possible, is to step back and track the changes as they come in, proactively considering the changes, just like we do for regular source code, which allows for far more robust change management and accountability for the database.

So that’s the idea.  In the next few weeks I’ll have some more posts on this topic, getting into some more specifics about how to solve this.  Hopefully we can find a simple, easy to manage solution, that developers can adapt to quickly, and which promotes database changes to the first-class citizen of the change management world that it deserves to be.