Data modeling is in trouble. But perhaps not for the reasons you might suspect.
That’s the clearest signal from the Practical Data Community 2026 State of Data Engineering Survey. We asked 1,101 data professionals about their tools, challenges, and outlook. The section on data modeling stopped me cold. 89% report at least one pain point with their modeling approach. Only 11% say modeling is going well.
Is this a technology or skills problem? Not according to the data. The top complaints are organizational: pressure to move fast (59%) and lack of clear ownership (51%). Data modeling is failing because organizations won’t give it the time and attention it requires.
TL;DR - it’s your boss’s fault ;)
Is all of this dire? No. As we’ll see, there’s also a (difficult) path forward.
Modern data modelers need to live at the intersection of business and tech. Ellie.ai allows you to collaborate effectively with the business while maintaining credibility with the Tech team. Get contextual support from AI, reverse-engineer anything, build a repository of sources with synthetic AI-generated contextual metadata, deliver insights via an MCP Server, and integrate with anything via full-blown API support.
The full-stack data modeling future is here today! Learn more at Ellie.ai.
Thanks to Ellie.ai for partnering on this newsletter.
What People Actually Do Today
Before we get to the pain, let’s look at what teams are doing today.
Approach / %
Mixed (depends on use case) 37%
Kimball-style dimensional 28%
Ad-hoc / tables as needed 17%
Canonical / semantic models 5%
One Big Table 4%
Event-driven 3%
Data Vault 3%
“Mixed” leading the pack isn’t surprising. Most mature teams adapt their approach to the use case. What’s concerning is that 17% doing ad-hoc modeling. That’s roughly 1 in 6 teams adding tables as needed with no consistent methodology.
And only 5% have adopted semantic or canonical models, despite strong interest in the topic. More on that later.
The Pain Is Real
We asked respondents to select their data modeling pain points. They could choose multiple options.
Pain Point / %
Pressure to move fast 59%
Lack of clear ownership 51%
Hard to maintain over time 39%
Tools don’t support good modeling 19%
None / modeling is going well 11%
AI tools produce inconsistent schemas 4%
Read that top line again. Nearly 6 in 10 data professionals say pressure to move fast is their biggest modeling challenge. To me, this seems like a leadership and prioritization problem. It’s a “we need this dashboard by Friday” problem.
The second complaint - lack of clear ownership - reinforces this. When nobody owns the model, everybody adds to it, and nobody maintains it. I’m guessing the third reason - hard to maintain over time - is a result of the first two. From my anecdotal observations, these numbers are the norm.
The Firefighting Tax
“There’s never enough time to do it right, but there’s always enough time to do it over.” -Jack Bergman
Here’s where it gets interesting. We asked teams where they spend their time, and the top response was “fighting fires.” Overall, 26% of respondents said firefighting consumes significant time.
But when you break it down by modeling approach, a pattern emerges:
Modeling Approach / % Fighting Fires
Ad-hoc / tables as needed 38%
One Big Table 31%
Data Vault 31%
Kimball-style dimensional 24%
Mixed 23%
Event-driven 19%
Canonical / semantic models 19%
Ad-hoc modelers fight fires at twice the rate as teams using semantic models. This might also be a sample size issue, and something I’d like to dig into a bit more in upcoming pulse surveys.
This is the cost of skipping data modeling. You save time upfront. You pay for it later in the form of incidents, confusion, and rework. The 38% vs. 19% gap represents hundreds of hours of lost productivity per year.
Education and Training
Only 5% of respondents use canonical or semantic models today. But the results told a different story when it comes to training topics people are interested in:
Training Topic / # of Responses
AI/LLM integration 235
Data modeling 211
Semantics / ontologies / knowledge graphs 209
Semantic layer training is the third-most-requested topic, nearly tied with data modeling itself. There’s a massive gap between interest and adoption.
Why? My guess - AI is making semantics/ontologies/knowledge a popular topic. Sadly, there’s very little approachable training on the market. Hopefully, that changes. Also, semantic layers require upfront investment that’s hard to justify when you’re under pressure to move fast. The 59% who feel pressure aren’t going to pause for a six-month semantic modeling initiative.
But as the teams that do make that investment fight fewer fires. The data is clear on this.
The survey paints a picture of an industry that knows it has a modeling problem but can’t find the space to fix it. The pressure to move fast leads to ad-hoc models. Ad-hoc models create a maintenance burden. Maintenance burden creates firefighting. Firefighting consumes time that could be devoted to better modeling. The cycle continues.
Breaking this cycle requires leadership buy-in. It requires treating data modeling as infrastructure rather than a project. It requires someone with authority to say, “We’re going to slow down now so we can speed up later.”
The 11% who say modeling is going well? I’m guessing they figured this out. They just made modeling a priority before it became a crisis.
The Path Forward
If you’re in the 89% experiencing modeling pain, here’s what the data suggests:
Acknowledge it’s not a tooling problem. Only 19% blame their tools. The other 81% point to organizational issues. New software won’t fix unclear ownership or pressure to ship.
Quantify the firefighting tax. Track how much time your team spends on incidents, data fixes, and “why doesn’t this number match” investigations. Make the cost visible to leadership.
Start with ownership. Before you pick a methodology, pick an owner. Someone who can say no to ad-hoc table requests. Someone accountable for model quality over time.
Consider semantic models. The 5% using them fight fires half as often as the ad-hoc crowd. That’s not a coincidence. The upfront investment pays off.
Make the case for slowing down. The survey shows that teams under pressure to move fast end up spending more time firefighting. Speed without modeling discipline is an illusion.
One More Thing. A Program, Not a Project
We asked respondents what they wish the wider industry understood about data engineering. One response stuck with me:
“It’s not a project. It’s a program that needs to be treated as a capital investment.”
Data modeling is the same. It’s not something you finish. It’s something you maintain. The teams that understand this - the 11% - are the ones who say it’s going well.
Full survey results here.
Practical Data Modeling is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.