ClickBench

ClickBench is an open-source benchmark created by the ClickHouse team to compare analytical (OLAP) database engines on a common, realistic workload. It has become a widely referenced public leaderboard covering dozens of systems — columnar databases, data warehouses, query engines, and even some general-purpose databases — all running the same queries against the same data.

What It Measures

ClickBench measures single-query latency for 43 typical analytical queries over one large, denormalized web-traffic table of roughly 100 million rows. It records both cold (first run, no cache) and hot (repeated run) times, the data load time, and the on-disk storage size after loading. Together these capture raw query speed, caching behavior, ingestion cost, and compression efficiency.

Methodology

The data set is a single flat table of anonymized web-analytics hits with around 100 columns, distributed as compressed files. Each engine has a runner script that loads the data and executes the 43 queries — filters, aggregations, group-bys, top-N, and approximate-distinct operations typical of dashboards and ad-hoc analytics. Every query is run three times; the benchmark reports the minimum of the three (hot) and notes cold behavior. Results are normalized and published as a public table where systems can be sorted and compared, with the test harness and queries fully open so anyone can reproduce or add a system.

How to Interpret Results

Look at the full profile, not a single number: an engine may have fast hot queries but slow cold queries, large storage, or long load times. The single-table design means ClickBench rewards columnar storage, vectorized execution, and compression, and does not test joins. Because submissions come from many contributors on varied hardware, normalize for instance type before comparing, and prefer results you reproduce yourself for a purchasing decision. Hot times reflect cached performance; cold times better reflect first-touch analytics.

Limitations

The workload is a single denormalized table, so it ignores joins, which are central to real warehouse schemas like TPC-DS. It targets one web-analytics pattern and modest scale (about 100 million rows), so it may not predict behavior on multi-terabyte, multi-table workloads. Self-reported results across heterogeneous hardware require care. Use ClickBench to gauge single-table scan-and-aggregate speed, compression, and load cost, complemented by join-heavy benchmarks for full coverage. Its open, reproducible harness and broad coverage have made ClickBench a popular community reference for tracking the rapidly evolving field of analytical query engines. When evaluating an engine, normalize the published numbers to a common instance type or, better, run the open harness yourself on hardware that matches your intended deployment.