lbench - Benchmarking Tool for Python Projects

lbench is a benchmarking tool built on top of pytest and pytest-benchmark, designed to make it easy to write, run, and analyze benchmarks for Python projects. It provides automatic result logging, cProfile profiling and flamegraphs, Dask performance reporting, memory tracking, a Jupyter notebook magic, and a dashboard for visualizing and comparing benchmark results over time.

Installation

git clone https://github.com/yourusername/lsdb-benchmarking.git
cd lsdb-benchmarking
pip install -e .

Writing Benchmarks

Basic Benchmarks with lbench Fixture

The lbench fixture extends pytest-benchmark with automatic cProfile profiling:

def test_my_function(lbench):
    def benchmark_func():
        result = my_function()

    lbench(benchmark_func)

Add @pytest.mark.lbench_memory to also track peak memory usage with memray:

@pytest.mark.lbench_memory
def test_my_function(lbench):
    lbench(my_function)

Dask Benchmarks with lbench_dask Fixture

def test_my_dask_function(lbench_dask):
    def benchmark_func():
        result = my_dask_dataframe.compute()

    lbench_dask(benchmark_func)

The lbench_dask fixture automatically collects Dask task stream information, generates a Dask performance report, and samples memory usage during execution. Use lbench_dask_collection to also record the Dask graph size and node count:

def test_collection(lbench_dask_collection):
    catalog = lsdb.read_hats(...)
    lbench_dask_collection(catalog)

Running Benchmarks

pytest --lbench benchmarks/

This creates a timestamped result directory, runs all benchmarks, and saves:

• pytest-benchmark.json — timing stats and extra metrics
• cprofile_*.prof — cProfile data for each benchmark
• dask_performance_report_*.html — Dask performance reports (Dask benchmarks only)
• memray_*.bin — memory tracking data (memory-marked benchmarks only)

Configuring the Result Directory

# Via flag
pytest --lbench --lbench-root=/path/to/results benchmarks/

# Via environment variable
export LBENCH_ROOT=/path/to/results
pytest --lbench benchmarks/

If neither is set, results are saved to ./lbench_results.

Running Benchmarks in Jupyter Notebooks

lbench provides a %%lbench cell magic that produces the same JSON log format as the pytest runner, so notebook results appear alongside pytest results in the dashboard.

Load the extension once per notebook:

%load_ext
lbench.notebook

Then use the cell magic on any cell:

%%lbench
my_expensive_function()

With options:

%%lbench - -rounds
10 - -warmup - -memory - -profile - -name
my_benchmark
my_expensive_function()

Available options:

Dask benchmarks in notebooks

%%lbench - -dask - -rounds
3
my_collection.compute()

# With graph stats from a named variable:
%%lbench - -dask - -collection
src_catalog - -name
catalog_scan
src_catalog.compute()

Results within a notebook session are accumulated into a single timestamped run directory. Call lbench.notebook.magic.reset_session() to start a fresh run directory mid-session.

Viewing Results with the Dashboard

lbench dash
lbench dash --port 8051

Or from a notebook:

from lbench.dashboard.app import run_dashboard

run_dashboard(port=8050)

Calling run_dashboard() again will restart the server on the new settings.

lbench dash and run_dashboard() use the Dash development server. To host the dashboard, point a WSGI server at lbench.dashboard.app:server instead:

pip install gunicorn
LBENCH_ROOT=/path/to/results gunicorn --bind 0.0.0.0:8050 lbench.dashboard.app:server

Dashboard Features

Run browser (sidebar)

• Lists all runs in chronological order
• Filter runs by date range with the date picker
• Rename runs with the pencil icon (temporarily disabled for security)

Benchmark tables

• Per-benchmark cards showing timing stats, memory usage, and Dask metrics
• Links to open flamegraphs (cProfile) and Dask performance reports directly in the browser

Trend plots

• Click "Plot series" to switch to the trend view
• Select one or more benchmarks and a metric to plot performance over time
• Error bars show standard deviation where available
• Respects the active date filter

Example

import pytest


@pytest.mark.parametrize("size", [1000, 10000, 100000])
def test_dataframe_operation(size, lbench):
    import pandas as pd

    df = pd.DataFrame({'A': range(size), 'B': range(size)})

    def benchmark_func():
        result = df['A'] + df['B']

    lbench(benchmark_func)

pytest --lbench tests/test_dataframe_operation.py