Nikola Ilic shows off a relatively new feature:
The first thing every Fabric architect reaches for in this situation is the usual playlist: check the query plan, look at the joins, validate the statistics, maybe scale up the capacity. All worth doing, but none of those things addressed what was actually happening: the warehouse was scanning the entire table for every filtered query, because there was no way to tell it which Parquet files actually contained the rows we cared about.
However, Microsoft shipped data clustering in preview at the end of November 2025, and the entire conversation changed.
In this article, I want to walk you through what data clustering is, how it works under the hood, and most importantly, I’ll show you a real demo on a 100-million-row clickstream table that you can run in your own warehouse. No abstractions, no marketing numbers, but actual T-SQL you can paste.
Some of the notes Nikola mentions remind me of some of the rules around making columnstore indexes work and for much of the same reason. But as Nikola’s demo shows, this is definitely a “You must be this tall to ride the ride” feature, and unless you’re talking about quite large fact tables with (at a minimum) billions of rows of data, the benefit mostly comes from reducing CUs rather than wall clock time improvements.
Comments closed