VLDB 2026 — Experiment, Analysis & Benchmark track

Redbench

Workload Synthesis From Cloud Traces

Johannes Wehrstein¹, Roman Heinrich^1,2, Mihail Stoian³, Skander Krid⁴, Martin Stemmer¹, Andreas Kipf³, Carsten Binnig^1,2,6, Muhammad El-Hindi⁵

¹TU Darmstadt ²DFKI ³UTN ⁴Snowflake ⁵TU Munich ⁶hessian.AI

Real cloud traces look nothing like TPC-H or TPC-DS. Redbench turns them into runnable SQL workloads that keep what standard benchmarks throw away — query repetition, arrival times, and a realistic read/write mix.

Paper ↗ Code ↗ Workloads ↗ Evaluation ↗

About

What is Redbench?

Cloud providers have published traces of what their data warehouses actually run. Redbench reads those traces and rebuilds their characteristics on top of a benchmark or schema you already have, turning a static query set into a realistic query stream. It keys on inherent workload signals rather than execution metrics, so the result is reproducible across systems.

Trace source: RedsetAmazon Redshift metadata, 200 instances
Query types: Read & writeSELECT, INSERT, UPDATE, DELETE
Evaluated on: 4 systemscommercial cloud data warehouses

Synthesis

Two ways to build a workload

Run either strategy, or both in sequence — they answer different questions.

Matching

Reuse queries you already trust

Maps each trace query onto one from an existing benchmark — TPC-DS, JOB, CEB — by join count and scanset, keeping semantics you already understand.

Generation

Synthesize queries for your schema

Walks a join graph built from your schema and tunes predicate selectivity to hit the trace's bytes-read distribution. Works on any database you supply.

Criterion	Matching	Generation
Query types	SELECT / simple DML	SELECT / INSERT / UPDATE / DELETE
Complexity	Complex, follows TPC-H / JOB	Many joins, simple predicates
Repetition	Exact	Structural & exact
Matches Redset on	Arrival times, schema, joins	Arrival times, schema, joins, bytes read

Quickstart

Generate a workload

Python via uv, plus the DuckDB CLI. Configuration and custom schemas are covered in the README.

git clone https://github.com/DataManagementLab/Redbench.git
cd Redbench && uv sync && source .venv/bin/activate

# fetch a Redset trace — serverless is the small one
mkdir data
aws s3 cp --no-sign-request \
  s3://redshift-downloads/redset/serverless/full.parquet data/full_serverless.parquet

# synthesize
python src/redbench/run.py --redset_path data/full_serverless.parquet --output_dir output

Or skip generation and download the paper's workloads from OSF.

Citation

Cite Redbench

@article{wehrstein2025redbench,
  title   = {Redbench: Workload Synthesis From Cloud Traces},
  author  = {Wehrstein, Johannes and Heinrich, Roman and Stoian, Mihail and
             Krid, Skander and Stemmer, Martin and Kipf, Andreas and
             Binnig, Carsten and El-Hindi, Muhammad},
  journal = {arXiv preprint arXiv:2511.13059},
  year    = {2025}
}