Re: Power BI September 2026 Feature Summary

Hi tate-bowman​, your comment made me curious, so I tested it (F64 trial, Direct Lake on OneLake, 100M rows, 36.7M distinct customers).

You're partly right. With the column already in memory, APPROXIMATEDISTINCTCOUNT is still much faster on warm queries: 8.6 s -> 0.3 s for a single total, 28 s -> 0.7 s grouped by day, and about 10x fewer CU across the same 60 queries. On a cold start the advantage drops to 1.3-2x, because loading the column takes about 20 s either way.

On memory you're right. Each ADC query used at least ~1 GB, and with 150k small groups (~550 distinct customers each) it hit the 10 GB per-query limit, while DISTINCTCOUNT handled the same query in 3.3 GB. So it doesn't help with data that doesn't fit in memory - pre-aggregated, mergeable HLL sketches as in your idea would still be the way to go there.

One more thing for the product team: when Direct Lake on SQL falls back to DirectQuery, the measure fails with "isn't supported for this data source", even though the SQL analytics endpoint itself supports APPROX_COUNT_DISTINCT.

Full write-up (in Polish): https://lnkd.in/p/eadzuM-2