Semantic Layer Middleware is Not Enough

Even Cube (formerly CubeJS) has pivoted to adding Workbooks, Dashboards,and Chat interfaces

In a not so surprising turn of events one of the often referenced semantic layer as middleware companies has pivoted to providing actually useful interfaces. Cube now has workbooks, dashboards, and AI chat interfaces. All the parts of a true full stack business intelligence platform. How long before they drop support for “connect all your BI tools” nonsense?

In reality they may keep all of those connectors nominally. But it will just wither away with time. How could you cause all your connecting BI tool to leverage the semantic layer as effectively as your own UI? You can’t. The UI will impose its own demands on the semantic layer and vice a versa. Those UI driven improvements will only be accretive to Cube UI surfaces. A good wedge strategy to eventually replace their connecting “partners”.

This should be a warning to every Enterprise deploying semantic middlewares. Whether its Cube, MetricFlow, Semantic Views (snowflake), homegrown, or something else you should know this is not enough. This middleware strategy is fundamentally flawed if your goal is to enable broad self service for humans and safe data retrieval for agents. The problem is architectural. In middleware strategies tables and table like structures are the primitive. Every interface needs to be conversant at this level.

In our opinion this is too low level for the primary use case of self service. The optimal level is the semantics themselves: measures and dimensions.

There is no table.

Tying everything to the table level is just legacy thinking. My buddy coined the term “BI brain”. Sort of a derogatory reference to folks on our team who could only think of building software applications with ETL and data warehouse concepts in the mix. They just can’t think beyond that anchor point. This is the sort of thinking thats getting us “Semantic Views”.

Interestingly, Cube is not the first to pivot out of middleware only strategy. DremIO started of as a middleware solution like Semantic Views. They eventually pivoted to full lakehouse platform with some semantic stuff. Talk about pivots! And before this, in eras past, data virtualization with federated warehousing software was all the rage. You don’t need a data warehouse just user our virtualization software! Many of these companies are now defunct or pivoted beyond recognition. In their wake lie hours of wasted human capital. I believe semantic layer middleware trend may be headed in that direction.

The only saving grace and possible utility of a semantic middleware is its potential to enable better coordination amongst data engineers. Your base metrics are defined in some construct from which you can quickly build a new pipeline to derive some new metric.

However, your an engineer with Claude, how much value is this construct really adding?