Semantic Layers: The Right Idea, the Wrong House - Issue 264
The case against semantic layers (until you actually need one)
Welcome to the Data Analysis Journal, a weekly newsletter about data science and analytics.
“If you can't dazzle them with brilliance, baffle them with bullshit.”
I don’t know who said that or why, but it fits as an opening to my first-ever introduction to the semantic layers in my newsletter.
My readers know I’m skeptical of most hyped and bestselling concepts in data, such as CDPs (How To Make A Sandwich in 587 Steps), data quality products (Does data quality have ROI?), self-serve BI (When Things Go South), or old-school BI (Tableau vs. Power BI – Which Should You Choose? Neither.) Many of these concepts and solutions are “good on paper,” but in practice, they overcomplicate the stack, overwhelm teams, create more “data noise,” and add exponential cost without delivering meaningful value.
I did my best to avoid semantic layers for a very long time, because, unlike the examples above, this isn’t the usual straightforward BS we’ve been sold. While I’d personally prefer not to have any layers “in my house,” based on what I see happening in other “houses,” I’d say teams are way safer with semantic layers than without.
In this publication, I’ll do my best to democratize data layering and cover:
The difference between semantic layers and metrics layers (these are not the same!). When you should use what and why.
Managing the trade-off between “get it now” and “get it right”
How to reach the dreamed select * from analytics.metrics.
Helping you decide when it’s right for your team to implement a semantic layer - or if you’re better off without it.
Semantic layers ≠ Metrics layers
This is probably the most misunderstood concept across data leadership. Many think of semantic layers as aggregated layers of metrics with pre-defined logic and formulas to be consumed by BI.
Not quite.
And the reason for such confusion is that the definition of a semantic layer often depends on the particular data stack being used. For example, Power BI has a built-in semantic layer that provides a layer of security, defines metrics, and implements hierarchies. In this case, the semantic layer and the metrics layer are combined. However, if you're not using Microsoft products, other systems, such as dbt, Snowflake, or Looker, differentiate between the transformative layer and the metrics layer and offer different tools (products?) for working with them.
So yes, both semantic and metrics are layers. Both sit between the ingested data and BI. And both are meant to make raw data consumable for BI in the shortest time with the highest value. But in different ways.
