Challenging the Parquet Dogma: CSV vs Parquet Benchmark

“Tested like an engineer, visualized like a designer, reported like a professional.”

🚀 Project Overview

This repository rigorously benchmarks the performance of CSV and Parquet file formats using Polars in Python. It challenges common assumptions that Parquet universally outperforms CSV in all use cases. By analyzing file loading speed, memory efficiency, and runtime profiling under both cold (after reboot) and warm (cached) OS-level conditions, this project provides real-world, reproducible insights for data engineers and analysts.

---

🎯 Key Features & Technologies

• Benchmarking: Compare CSV and Parquet file formats for local analytics workflows
• Polars: Ultra-fast DataFrame operations (Rust backend)
• psutil: Precise per-process memory measurement
• cProfile & SnakeViz: In-depth runtime profiling and flame graph visualization
• Matplotlib: Automated result charting
• Cross-platform: Works on Windows, macOS, and Linux

---

📂 Project Structure

benchmark_project/
├── benchmark.py               # Main benchmarking script
├── analyze_profile.py         # Summarize cProfile results, generate images
├── generate_profile_images.py # Batch conversion to PNG via gprof2dot
├── snakeviz_viewer_all.py     # Opens all profiles in browser via SnakeViz
├── visualize_results.py       # Plots time/memory graphs from logs
├── requirements.txt           # Required Python packages
├── results/                   # Benchmark logs, .prof files, profile images (auto-generated)
├── sample_data/               # (Optional) Example CSV/Parquet for demo (not included by default)
└── README.md

docs/
├── challenging-the-parquet-dogma-tech-report.pdf   # Full technical report (PDF)
├── challenging-the-parquet-dogma-tech-report.pages # Apple Pages editable source

---

⚙️ Setup & Installation

1. Clone this repository

git clone https://github.com/its-spark-dev/csv-vs-parquet-benchmark.git
cd csv-vs-parquet-benchmark/benchmark_project

2. (Recommended) Create a virtual environment:

python -m venv venv
source venv/bin/activate # On Unix/macOS
venv\Scripts\activate    # On Windows

3. Install dependencies:

pip install -r requirements.txt

---

🚦 How to Run the Benchmark

1. Prepare Data

• By default, the script looks for csv_data/ and parquet_data/ in the project folder.

• These data folders are not included in the repo (to save space).

  • Use your own test files, or generate synthetic data as needed.
  • Optionally, add a few small sample files to sample_data/ for quick demos.

2. Run the Benchmark

# Simulate cold cache (run after reboot or first run)
python benchmark.py --cache cold

# Simulate warm cache (repeat run)
python benchmark.py --cache warm

• Results (logs, .prof, .png) are written to results/.

3. Visualize Results

Charts:

python visualize_results.py

Flame Graphs (SnakeViz):

python snakeviz_viewer_all.py

Static Profile Images:

python analyze_profile.py

---

📊 Sample Results (from actual runs)

Cache Type	Format	Avg Time (sec)	Avg Memory (MB)
Cold	CSV	0.29	32.6
Cold	Parquet	0.82	205.1
Warm	CSV	0.25	14.8
Warm	Parquet	0.82	132.9

Note: Despite Parquet’s reputation, CSV outperformed it for small files in these experiments. Profiling details and rationale are below.

---

🔥 Figure: SnakeViz Flame Graph Example

Figure. Flame graph showing the call stack for reading Parquet files. Most time is spent in collect() during LazyFrame execution.

---

🔬 Profiling Analysis (Why CSV Was Faster)

• Polars loads Parquet via lazy evaluation, so .collect() triggers full data materialization — adds runtime overhead.
• CSV is read eagerly; no equivalent .collect() step.
• For many small files, CSV’s simplicity and reduced overhead produced superior performance in this context.
• See the included technical report and flame graph images for deep-dive analysis.

---

📌 When Parquet Is Still Better

Parquet still excels for:

• Large-scale, columnar queries
• Distributed/cloud-based data processing
• Selective I/O or columnar reads

CSV remains a better choice for:

• Simple, local batch jobs
• Lightweight analytics
• Immediate data access (no extra tooling)

---

🧪 Benchmark Environment

• OS: Windows 10 Pro, Version 22H2, Build 19045.5854
• Python: 3.13.3 (64-bit)
• CPU: AMD Ryzen 5 5600X — 6C/12T @ 3.70 GHz
• RAM: 15.9 GB
• GPU: NVIDIA GeForce RTX 3060

---

🗂️ Data Policy

Note: Large-scale experimental data files (csv_data/, parquet_data/) are not included. To reproduce experiments, generate your own sample files or contact the maintainer. Example scripts or instructions can be provided.

---

📄 Full Technical Report (Downloads)

• 📄 Download PDF: Challenging the Parquet Dogma – Tech Report

• 📝 Download Apple Pages (editable): Challenging the Parquet Dogma – Tech Report

  • The full report includes background, benchmarking methodology, results, profiling, and recommendations for real-world file format selection.

---

🙋 Maintainer & Contact

Sanghyeon Park

• LinkedIn
• GitHub

---

📝 License

MIT License

---

“Tested like an engineer, visualized like a designer, reported like a professional.”