RepoRoulette 🎲: Randomly Sample Repositories from GitHub

Spin the wheel and see which GitHub repositories you get!

🚀 Installation

# Using pip
pip install reporoulette

# From source
git clone https://github.com/gojiplus/reporoulette.git
cd reporoulette
pip install -e .

📖 Sampling Methods

RepoRoulette provides three distinct methods for random GitHub repository sampling:

What population does each sampler draw from?

The four samplers draw from different populations, and only one is (approximately) uniform over all repositories. Pick the one whose population matches your research question:

• IDSampler — ≈ uniform over all existing public repositories with ID ≤ max_id. The closest thing to a true random sample of GitHub; the cost is a moderate hit rate (~37% measured live in 2026; the rest are deleted, private, or unassigned IDs). The default max_id is a dated constant — call update_max_id() (one API call) to cover the newest repositories.
• TemporalSampler — repositories pushed on randomly chosen days. A repository active on many days has proportionally more chances of being sampled, so the sample is biased toward actively maintained projects. The GitHub Search API also caps results at 1,000 per query, so on high-activity days only the most recently updated 1,000 repositories are reachable.
• BigQuerySampler — repositories generating GH Archive events on sampled days. Event-based, so also biased toward active repositories.
• GHArchiveSampler — with the default CreateEvent filter, repositories created on sampled days; with other event types, an activity-biased event population like the BigQuery sampler. Two caveats: (1) GitHub's Events API change of 2025-10-07 removed repository-creation events from the public feed, so the default population is empty for days after that date — no repository created since then is reachable; (2) GH Archive publishes hourly files, and hours_per_day trades bandwidth for population: the default (all 24 hours) samples the true day population at ~2 GB/day, while hours_per_day=H downloads only H files but redefines the population to repositories active in the sampled hours (over-representing low-traffic hours, the same bias structure as the per-day cap).

1. 🎯 ID-Based Sampling

Uses GitHub's sequential repository ID system to generate truly random samples by probing random IDs from the valid ID range. The downside of using the method is that the hit rate can be low (as many IDs are invalid, partly because the repo. is private or abandoned, etc.) And any filtering on repo. characteristics must wait till you have the names.

The function will continue to sample till either max_attempts or till n_samples. You can pass the seed for reproducibility.

from reporoulette import IDSampler

# Initialize the sampler
sampler = IDSampler(token="your_github_token")

# Get 50 random repositories
repos = sampler.sample(n_samples=50)

# Print basic stats
print(f"Success rate: {sampler.success_rate:.2f}%")
print(f"Samples collected: {len(repos)}")

2. ⏱️ Temporal Sampling

Randomly selects days within a specified date range and retrieves repositories updated during those periods using weighted sampling based on repository activity.

from reporoulette import TemporalSampler
from datetime import datetime, timedelta

# Define a date range (last 3 months)
end_date = datetime.now()
start_date = end_date - timedelta(days=90)

# Initialize the sampler
sampler = TemporalSampler(
    token="your_github_token", start_date=start_date, end_date=end_date
)

# Get 100 random repositories
repos = sampler.sample(n_samples=100)

# Get repositories with specific characteristics
filtered_repos = sampler.sample(
    n_samples=50,
    min_stars=10,
    language="python",  # Note: single language, not list
)

3. 🔍 BigQuery Sampling

The BigQuerySampler leverages Google BigQuery's public GitHub dataset to sample repositories with advanced filtering capabilities.

Setup for BigQuery Sampler

1. Create a Google Cloud Platform (GCP) project:

   • Go to the GCP Console
   • Create a new project

2. Enable the BigQuery API:

   • In your project, go to "APIs & Services" > "Library"
   • Search for "BigQuery API" and enable it

3. Create a service account:

   • Go to "IAM & Admin" > "Service Accounts"
   • Create a new service account
   • Grant it the "BigQuery User" role
   • Create and download a JSON key file

4. Install required dependencies:

   pip install google-cloud-bigquery google-auth
   

5. Using BigQuerySampler:

from reporoulette import BigQuerySampler

# Initialize with service account credentials
sampler = BigQuerySampler(
    credentials_path="path/to/your-service-account-key.json",
    project_id="your-gcp-project-id",
    seed=42,
)

# Sample active repositories with commits in the last year
active_repos = sampler.sample(
    n_samples=50,
    population="active",
    languages=["Python", "JavaScript"],  # Optional language filter
)

# Sample repositories across random days
random_repos = sampler.sample_by_day(n_samples=50, days_to_sample=10, years_back=5)

# Get language information for sampled repositories
languages = sampler.get_languages(random_repos)

# Print results
for repo in random_repos:
    print(f"Repository: {repo['full_name']}")
    repo_languages = languages.get(repo["full_name"], [])
    if repo_languages:
        print(f"Primary language: {repo_languages[0]['language']}")
    print("---")

Advantages:

• Handles large sample sizes efficiently
• Powerful filtering and stratification options
• Not limited by GitHub API rate limits
• Access to historical data

Limitations:

• Could be expensive
• Requires Google Cloud Platform account and billing
• Dataset may have a slight delay (typically 24-48 hours)

4. GH Archive Sampler

The GHArchiveSampler fetches repositories by sampling events from GitHub Archive, a project that records the public GitHub timeline.

from reporoulette import GHArchiveSampler

# Initialize with optional parameters
sampler = GHArchiveSampler(seed=42)  # Set seed for reproducibility

# Sample repositories
repos = sampler.sample(
    n_samples=100,  # Number of repositories to sample
    days_to_sample=5,  # Number of random days to sample from
    repos_per_day=20,  # Repositories to sample per day
    years_back=2,  # How many years to look back
    event_types=[
        "PushEvent",
        "CreateEvent",
        "PullRequestEvent",
    ],  # Event types to consider
)

# Access results
for repo in repos:
    print(f"Repository: {repo['full_name']}")
    print(f"Event Type: {repo['event_type']}")
    print(f"Sampled From: {repo['sampled_from']}")
    print("---")

📊 Example Use Cases

• Academic Research: Study coding practices across different languages and communities
• Learning Resources: Discover diverse code examples for education
• Data Science: Build datasets for machine learning models about code patterns
• Trend Analysis: Identify emerging technologies and practices
• Security Research: Find vulnerability patterns across repository types

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

🔗 Related Projects

• GHTorrent - GitHub data archive project
• GitHub Archive - Archive of public GitHub timeline
• PyGithub - Python library for the GitHub API

---

Built with ❤️ by Gojiplus