Read the full article: Going Raw Dog on the Database -- After 25+ years of championing ORMs, Michael Kennedy makes the case for abandoning them in favor of raw database queries paired with Python dataclasses. The Raw+DC pattern delivers better AI coding assistance, fewer dependencies, and comparable or superior performance -- all while keeping type safety at your data access boundaries. This repo contains the benchmark code behind those findings.

A performance benchmark suite comparing four Python approaches to MongoDB:

Read benchmarks (per library):

• Single-field projection (category name, order email)
• Full document fetch (simple category, complex order with nested subdocs)
• Bulk reads: 100 / 1,000 / 10,000 orders by status
• Bulk reads: 100 / 1,000 / 10,000 categories sorted by view count

Write benchmarks (per library):

• Single insert
• Batch insert (100 categories, 1,000 orders)
• Single update
• Single delete

Each benchmark runs 10 iterations and reports median, min, max, p95, and mean times in milliseconds.

• Python 3.13+
• MongoDB running on localhost:27017 (no auth)

python -m venv venv
source venv/bin/activate
uv pip install -e .

Populates categories and orders collections with 100K documents each, then creates indexes.

python main.py seed            # Seed (skips if already populated)
python main.py seed --force    # Drop and re-seed from scratch

python main.py run                         # Run everything
python main.py run --reads                 # Read benchmarks only
python main.py run --writes                # Write benchmarks only
python main.py run --library beanie        # Single library only
python main.py run --library raw --reads   # Combine filters
python main.py run --no-charts             # Skip chart generation

The run command auto-seeds if the database isn't populated.

python main.py reset    # Drop the entire benchmark database

Results are displayed as Rich tables in the terminal:

• Comparison table -- side-by-side timings per benchmark, fastest highlighted in green
• Overhead table -- multiplier showing how much slower each ODM is vs raw PyMongo

Charts are saved to output/:

• read_benchmarks.png -- grouped bar chart of read timings
• write_benchmarks.png -- grouped bar chart of write timings
• overhead_comparison.png -- overhead multiplier vs raw baseline

main.py                  # CLI entry point
config.py                # Constants (URI, DB name, iterations, seed count)
db.py                    # Connection helpers

models/                  # Data models for each approach
raw/                     # Raw PyMongo benchmarks (dicts)
dataclasses_raw/         # Dataclasses + PyMongo benchmarks
beanie_odm/              # Beanie async ODM benchmarks
mongoengine_odm/         # MongoEngine sync ODM benchmarks

benchmarks/              # Registry, runner, and timing utilities
seeding/                 # Data generation and database seeding
reporting/               # Rich tables and Matplotlib charts
output/                  # Generated chart PNGs