# ClickBench methodology

Reproduce analytical measurements with an exact public artifact, upstream queries and the declared cold/hot protocol.

Source: https://particledb.ai/benchmarks/clickbench/

ClickBench measures an analytical workload over one flat table. It does not establish transaction correctness, general application throughput or performance on every data model.

## Publication status

This page publishes methodology, not a leaderboard placement. Internal candidate measurements are not accepted upstream results. A ranking needs the dated comparison snapshot, hardware and filters; it can change as other entries change.

## Reproducible release

A submission must identify an immutable public binary URL and SHA-256, server version and edition, client version, upstream revision, operating system, hardware and storage. Publish the entry scripts, schema, all query definitions, raw results and failures. Installation on a fresh machine must not depend on a private source checkout, an existing binary or an unpublished SDK.

A result from one binary is not evidence for a different installer download. If the build or its optimization profile changes, repeat the measurements before replacing the reported artifact.

## Workload and storage

Declare the dataset, row count, load command and persisted on-disk size. Document storage-engine and collation settings. For a segment-table entry, state its supported operations explicitly; a read-only analytical result does not establish UPDATE/DELETE support or mixed transactional performance.

## Upstream measurement protocol

Follow the [ClickBench rules](https://github.com/ClickHouse/ClickBench/blob/main/README.md) at the submitted upstream revision:

- Run all 43 queries. Preserve the first cold run and the two following hot runs; do not add internal-suite warmups.
- A true cold run requires a server restart and OS page-cache clearing before each query's first run. Otherwise disclose the limitation using upstream's required tag.
- Disable query-result caching. Document other caching and configuration choices.
- Measure sending, executing and receiving results. Do not replace result delivery with a null-output mode.
- Keep failed or unavailable measurements visible as nulls; do not omit them to improve an aggregate.

Use the upstream aggregation rules. A normalized timing ratio above one means slower than its reference, not a speedup. Link the exact comparison view before quoting a score or placement.

## Independent reproduction

Have another developer run the public instructions on a fresh matching machine. Preserve correctness checks, binary hashes, full logs and deviations. This is an independent reproduction, not upstream acceptance or a guarantee that every repetition falls within a fixed percentage.

[Shared methodology](/benchmarks/overview/) · [Release availability](/installation/releases/) · [OLAP comparisons](/benchmarks/olap/)
