ClickHouse’s 752× Databricks claim is really a test of the fresh-data path

ClickHouse says its cloud service delivered 752× better end-to-end performance per dollar than a tested Databricks stack while continuously ingesting 113.2 billion stock-market quotes. The headline is striking, but the useful finding is narrower: this is a test of how cheaply each architecture turns arriving data into query-ready tables and summaries—not a universal database speed ranking. ClickHouse

What the benchmark actually ran

CostBench targeted one million rows per second while four aggregate queries ran every ten minutes and two drill-down queries ran hourly. ClickHouse used asynchronous inserts and incremental materialized views; Databricks used Zerobus Ingest, liquid clustering and triggered incremental materialized-view refreshes. The read side was matched at roughly 16 CPUs, and result caching was disabled. ClickHouse reports that Databricks completed durable ingestion in about 31.5 hours, versus about 36.75 hours for ClickHouse, so the claimed advantage did not come from finishing ingestion sooner. ClickHouse

The separation appeared after arrival. ClickHouse priced preparation at $28.69 and accepted query work at about $0.05, with 56.39 seconds of accumulated query runtime. Its Databricks model put preparation at $695.60, accepted query work at $2.65 and accumulated runtime at 29.10 minutes. ClickHouse multiplied combined preparation-plus-query cost by accumulated runtime to produce scores of 1,620.31 and 1,219,265.83—a ratio of 752.49. ClickHouse

Why the boundary matters

This is a vendor-run benchmark, and its score is not a standard industry metric. It combines two quantities—modeled cost and total query time—so a system that is both cheaper and faster compounds its lead. The cost model uses public list prices and allocated Databricks DBUs, not reconstructed invoices; it excludes storage, producer infrastructure, network charges, idle or minimum warehouse billing, and post-ingestion maintenance. The underlying stock-quotes dataset also requires a separate licence and cannot be redistributed, although ClickHouse published the harness, contracts and result artifacts. ClickHouse

The largest product caveat is explicit: the comparison used Databricks Serverless SQL because Lakehouse//RT, Databricks’ beta real-time warehouse, was unavailable in the test workspace. That means the result describes the tested production stack, not Databricks’ newest real-time architecture. ClickHouse

For practitioners, the defensible takeaway is architectural. Query latency alone can conceal the cost and lag of ingestion, clustering and summary refresh. A procurement test should therefore measure the complete fresh-data path, record when rows become query-visible—not merely durable—and verify summary freshness alongside query speed. ClickHouse’s 752× figure needs independent reproduction, but its benchmark contract is a better starting point than a dashboard-only speed claim.