# Benchmarks and methodology

Read benchmark results only with their workload, hardware, versions, durability settings and raw measurements.

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

Performance depends on the workload and configuration. A useful result includes enough information to reproduce both its behavior and its timing.

This documentation revision does not carry forward comparative headlines without a matching, reproducible result bundle. Application examples are correctness checks, not performance claims.

## Workloads

| Workload | What it measures | Methodology |
|---|---|---|
| Analytical queries | Load, cold and warm query behavior | [ClickBench](/benchmarks/clickbench/) |
| Operational transactions | A defined transaction mix under concurrency | [Transaction workloads](/benchmarks/tpc-c/) |
| Application analytics | Queries over representative operational data | [OLAP comparisons](/benchmarks/olap/) |
| Key-value | Durable SET, GET, INCR, list and hash throughput against Redis | [Key-value: ParticleDB and Redis](/benchmarks/key-value/) |

## Required result bundle

Every published run must identify:

- Server and client versions and the source revision where available.
- Hardware, operating system, CPU count, memory and storage.
- Dataset source, scale and query or transaction definitions.
- Indexes, transaction settings and durability settings.
- Warmups, timed repetitions and the reported statistic.
- Per-query or per-operation results, raw logs and failures.
- Competitor versions and matching settings for a comparison.

## Run discipline

Run one benchmark at a time on an otherwise idle machine. Follow the named suite's protocol: ClickBench uses its upstream cold/hot sequence without additional warmups; internal suites with two warmup iterations must retain them. Repeat internal comparisons and name the reported statistic. For before/after comparisons, run both versions on the same machine in the same session.

The 10-million-row cases generally give a more stable comparison than small samples. Larger cases can amplify variation from machine load.

## Compare like with like

Do not compare relaxed persistence with durable commits, cached data with a cold run, different query semantics, or failed work with successful work. Report the gaps as well as the favorable results.

## Current validation

Documentation and result-format checks do not establish a performance ranking. Measurements must be repeated in an idle environment before a new number is published.
