Rinil-Parmar/tpch-query-routing-dataset

Cross-Engine TPC-H Cost Modeling Dataset

📘 Overview

This dataset contains execution-time measurements and structural query features for SQL workloads executed across multiple database engines. It is designed for research in learned cost modeling, cross-engine optimization, and zero-shot SQL engine selection.

The dataset is generated using the TPC-H benchmark at Scale Factor .5 and includes automatically generated query variants for all 22 benchmark queries. Each query variant is executed on both SQLite and DuckDB, and execution times are recorded.

The dataset supports training machine learning models to predict query execution cost across heterogeneous database engines.

---

🎯 Purpose

Modern database systems increasingly support multiple SQL execution engines with different performance characteristics. However, selecting the most efficient engine for a given query remains a challenge.

This dataset enables:

• Cross-engine cost prediction
• Learned cost model research
• Engine-selection accuracy evaluation
• Zero-shot SQL engine routing
• Comparative performance analysis

---

🗂 Dataset Structure

Each row in the dataset corresponds to a single query variant.

Columns

Column	Description
query_id	TPC-H query identifier (Q1–Q22)
variant	Variant number of the query
sql	The executed SQL query
query_hash	Unique hash identifier of the query
sqlite_time_sec	Execution time on SQLite (seconds)
duckdb_time_sec	Execution time on DuckDB (seconds)
speedup_ratio	sqlite_time / duckdb_time
faster_engine	Engine with lower execution time
num_scans	Number of scan operators
num_joins	Number of JOIN operations
num_filters	Number of filter conditions
has_groupby	Whether query contains GROUP BY
has_orderby	Whether query contains ORDER BY
query_length	Length of SQL query string

---

📊 Dataset Statistics

• Total queries: ~1050
• Base queries: 22 (TPC-H)
• Variants per query: ~50
• Engines evaluated:
  • SQLite
  • DuckDB
• Scale Factor: .5
• Average runs per query: 3

---

🏗 Data Generation Process

1. TPC-H data generated using DuckDB’s tpch extension.
2. Same dataset loaded into both SQLite and DuckDB.
3. For each TPC-H query:
   • Random parameter variations generated.
   • Query executed multiple times per engine.
   • Execution time averaged.
4. Logical features extracted from SQL text.
5. Results stored in CSV format.

All experiments were executed in a controlled environment to reduce measurement noise.

---

🔬 Recommended Usage

This dataset can be used for:

• Regression modeling (execution time prediction)
• Classification (predict faster engine)
• Feature importance analysis
• Robustness and generalization experiments
• Cross-engine query optimization research

Example usage in Python:

import pandas as pd

df = pd.read_csv("master_dataset.csv")
X = df.drop(columns=["sqlite_time_sec", "duckdb_time_sec", "faster_engine"])