Timelake
Benchmarks

ClickBench: methodology

Planned

One wide table of web analytics, 100M rows; 43 queries.

Not run yet. We run a 10% subset during development. Comparable numbers need the full dataset on its reference machine. This page is the plan.

What we'll run

We'll use ClickBench's dataset, queries, and scripts, by its rules, for an official entry. Until then, a 10% subset runs in development and isn't published.

The original

ClickHouse's benchmark of analytical databases: 43 queries on a 100M-row table of web hits, each run cold and then twice hot, with load time and data size. Its public leaderboard includes DataFusion and DuckDB over Parquet and several data-lake setups.

How we'll run it

  • The full dataset, written by Timelake into a table on S3 and on local disk.
  • ClickBench's rules on its reference machine (c6a.4xlarge): a cold run and two hot runs per query, load time, size on disk.
  • Its queries and scripts are licensed CC BY-NC-SA, so they're downloaded at a pinned version rather than kept in our repository.
  • An official leaderboard entry once Timelake is more refined (issue #90).