Voice of the DBA

Category: Editorial

  • A Domain for Data

    A domain is a set of possible values (among other definitions). I use this word a lot in my work, often with a problem domain (the thing you’re trying to solve) or the domain of possible values (like the US States and Territories list). That last one is interesting, as this is often the set of data we stick in some reference or lookup table to use in a form on a screen.

    There is a domain as part of the SQL specification, which I never knew about. I was reading an article from Joe Celko on the CREATE DOMAIN statement. This doesn’t exist in SQL Server and T-SQL, but it has been a part of PostgreSQL for quite a few versions. The article talks about the definition of a domain from a few very experienced database design people.

    What’s interesting is that this could be a list of values, but it could be an expression against which values are checked. For example, we might choose the domain of positive integers, which might require that the value passes some x>=0 where x is the value. That seems OK, though this looks like a CHECK constraint to me, which we already have.

    The article is a little light and doesn’t give many examples of how or why a domain might be needed or why it might be better than a constraint. Perhaps there are good reasons, but I struggle to think of any. Certainly I am cautious of tightly binding which data is allowed, especially when I find many businesses have rules that often have exceptions.

    Would you use a domain for any data in your system? Do you think you have business rules that are firm and set enough to add domains? Perhaps you do, and if you do, that’s great. I suspect many of you are like me in that you are careful of where and how many check constraints you use. Those can be very hard to change and remove when an exception occurs, so the fewer the better.

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.

  • Database DevOps Metrics

    The DORA organization is dedicated to helping others build software better and faster, at a higher quality, and in a way that is more efficient. They continue to compile and publish the Accelerate State of DevOps report every year, which is a fascinating read.

    As a part of the report, they have identified four key metrics that identify high performing organizations in terms of software. These are divided into two areas: throughput and stability. Throughput measures are change lead time and deployment frequency. Stability measures are the change fail percentage and failed deployment recovery time.

    For a long time, as I chatted with various people doing database work, it seemed that most people deployed relatively infrequently. They might deploy a couple times a week for software changes, but database changes were often less than once a week. There have always been people moving faster or slower, but that felt like the pace for a majority of people. These days, in the 2024-2025 timeframe, many people seem to be able to deploy database changes every week, often multiple times a week.

    Lots of people have moved to more throughput, with more frequent deployments and less change lead time. Most of us can’t get more work out of people, so if we deploy more often, their completed work gets released quicker. Those two metrics make some sense, and I think those are good measures, but not goals. What I find is that people often need to make changes quicker either to respond to changing needs of their organization or to fix bugs they’ve introduced. I wonder what the ratio is of the former to the latter? I suspect it might be less than one, if most of your deployments are fixing bugs. I don’t mind deploying software quicker, but the design, modeling, and testing can’t be shortened.

    The stability metrics are often high for most people I speak with about deployments. I don’t see a lot of failures at deployment time as code usually compiles and deploys. It’s often a day (or week) later that someone notices the code doesn’t do what they expect. Is that a deployment failure? I think not. What’s the MTTR if it’s fixed an hour after being reported a day after the deployment? Is the MTTR an hour or a day plus an hour? I don’t know how these metrics apply to databases, especially if data gets mangled and has to be corrected manually over hours/days/weeks. Is that in the MTTR? Can you even track it?

    Metrics are good ways to measure you progress or health, as long as the metric doesn’t become the goal. I’ve run into a lot of customers using these metrics to measure their development, and it does help most for a period of time. Whether this continues to help them improve often depends on whether they keep focusing on their goals of delivering better quality software faster, or they focus on the metrics.

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.

  • Is GenAI Coming Faster Than You Think?

    If you’re a fan, no. You’re chomping at the bit, waiting to get everyone using GenAI models in their work. If you’re skeptical, then you might think it’s never coming to take your job, be a personal assistant, help you with coding, etc. Pick the task it won’t help you with.

    A more nuanced view, which is similar to mine, is from Kendra Little, in which she says AI will eliminate DBA Jobs  Faster Than You Think, I’m not quite as pro-AI as Kendra, mostly because I see so many companies that are slow to change, slow to adopt new tech, slow to adjust their thinking. They just soldier on and keep running their business, as they’ve been doing for decades. Even when you might make a case for change, or prove it’s worthwhile, they just don’t spend the time to change.

    Make no mistake, change takes time. And time is a valuable and limited resource. Even if you don’t appreciate time’s value, many of you still don’t want to spend time on new things.

    In the piece, Kendra notes that she has been using an AI Agent to help get work done. She sees that agent getting better at tasks and helping her get work done quicker. At some point, she knows the GenAI agent will be able to help her do the work of multiple people. Not eliminating DBA jobs, but reducing the need. Can we have 3 DBAs instead of 5 or 10? Can we get down to 2, or even 1 with a part-time DBA service?

    I do think that GenAI models and agents will help people get more work done, which might reduce the number of people an organization needs. However, I see no shortage of work in most organizations. In fact, I know of a few friends who can’t seem to hire a talented DBA. Perhaps a GenAI agent can support them and help them get work done without the need for a DBA. Not now, of course, but maybe in a year or two.

    However, humans still need to be in the process, and I suspect, even with an AI agent for every human, there’s still a lot of work to get done. I’m not convinced this will reduce employment. I think it could increase employment, though the bar for employees will rise.

    At least in some places. In some, they’ll keep doing the same thing they’ve been doing since 2001, or 1994, or 1983.

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.

  • The Inefficiencies of Kubernetes

    A report of cloud Kubernetes usage shows that these resources are being under-utiliized, over-provisioned, and costing more than necessary for many organizations. From the previous year, average CPU declined from 13% to 10%, and memory is used at only around 23%. Companies are over-provisioning their clusters, which is understandable. No one wants to have systems overloaded and users complaining about performance.

    However, this is a similar tension to what we see with virtualization on-premises. Operations people want to leave plenty of CPU/RAM/IO headroom for systems to handle bursting or increasing workloads. Management wants to get all the use they can out of their investment and would prefer we provision systems as closely as possible to their expected workloads. Containers and orchestrators should allow a closer match, but only if there are workloads that burst enough to require additional containers and pods to be deployed. That does happen with memory occasionally at a little over 5% of containers exceed their memory, but that’s not a significant amount.

    Managing a Kubernetes cluster is a specialized skill and most organizations don’t have the skills or experience to do it well. My view is that if you want to use an orchestrator, you’re better off letting the cloud providers manage the infrastructure and scale up and down as needed. There are autoscaling technologies to help Operations staff better manage their capacity and costs, but this is an additional skill people need.

    While I do think some companies are adopting cloud native technologies and rewriting their applications to run in containers and Kubernetes clusters, I find many more companies are hesitant to adopt a very complex technology on top of the complexity of teaching their developers to work within containers for their applications. Certainly in the Microsoft space, I don’t see a lot of database servers running in containers. Despite some of the advantages of upgrades and downgrades, the unfamiliarity with the ins and outs of containers leads most teams to continue to manage the database separately.

    Resource matching to a workload is a problem we’ve had for years and Kubernetes doesn’t make this any easier to deal with. The cloud is supposed to help us better manage our resources, but there is a lot of knowledge needed to do this well. Add in the cost/performance issues in the cloud and it’s no wonder that many companies have overprovisioned their resources to ensure systems continue running. I don’t know whether lots of IT staffers are optimistic about their workload growth or scared of potential problems from overloaded systems, but unless organizations carefully manage all their resources, they are likely to continue to see larger cloud bills than they like.

    Steve Jones

    Listen to the podcast at Libsyn, Spotify, or iTunes.

    Note, podcasts are only available for a limited time online.