Pull filtered slices of the GitHub Archive on a laptop, without holding raw hours on disk.
gharc is a command-line tool and Python library that filters the GitHub Archive dataset on consumer hardware. Each hourly archive is streamed through memory, filtered against your criteria, and written out as Parquet or JSONL. Peak local storage stays bounded by the downloads in flight at once, one temporary file per worker (each hourly archive is roughly 60 to 150 MB in 2024), so disk use scales with --workers rather than with how long a window you process.
The full GitHub Archive spans every public event since 2011: tens of terabytes compressed, and several petabytes uncompressed. Most studies need only a small slice of it, but the archive has no server-side filtering, so the usual options are downloading whole hours to local disk or querying a cloud-warehouse mirror (BigQuery, Snowflake).
gharc takes a third route and filters during the download:
- Streaming: Downloads each hourly archive (~60 to 150 MB compressed in 2024) to a temporary file.
- Filtering: Keeps only the events matching your filters (repositories, owners, actors, event types).
- Writing: Appends the matches to a single Parquet or JSONL output via
pyarrow.ParquetWriter. - Cleanup: Deletes the temporary download before moving to the next hour, so disk usage never accumulates.
It is meant for researchers and students who need event-level GitHub data over long windows but do not have a warehouse account or the disk to hold raw months.
- Bounded disk: Peak local storage is the downloads in flight, one temporary file per worker (about 250 MB at the default 4 workers, about 85 MB with a single worker), and it does not grow with the length of the window. For selective filters the working memory stays near 100 MB; a very wide or empty filter buffers more of each hour and uses more.
- Resumable downloads: Recovers from network interruptions (common on residential connections) using HTTP Range requests.
- Parallel processing: Hours in the range are downloaded and filtered across a thread pool.
- Filtering before parsing: A byte-level token check rejects lines that cannot match before any JSON parsing. With a selective filter most lines never reach the parser; a wide or empty filter gains nothing from this.
- Optional orjson: Uses
orjsonfor JSON parsing when it is installed, which is faster than the standard library parser. - Parquet output: Writes columnar data that loads directly into pandas, Polars, or Spark. JSONL output is available for resumable runs and line-oriented tools.
Measured on a Windows 11 laptop (12 logical cores, 15 GB RAM) over a typical residential connection. Reproducible scripts in benchmarks/.
A six-hour window of GHArchive (2024-01-01 00:00 to 06:00 UTC), filtered to apache/spark:
| Workers | Wall-clock | Hours/sec | Spark events | Peak RSS |
|---|---|---|---|---|
| 1 | 76.0 s | 0.079 | 14 | 94.2 MB |
| 4 | 58.1 s | 0.103 | 14 | 106.7 MB |
Both runs recovered the same events, so concurrency does not affect output. Peak RSS stays below 110 MB. The bottleneck on residential links is HTTPS download throughput rather than CPU; additional workers help up to a point and then saturate the connection.
Across the same six-hour window gharc streams about 416 MB of compressed source from GHArchive (six hourly files of roughly 60 to 85 MB each) but never retains it. The full source is still transferred, so this is not a bandwidth saving; what stays bounded is local disk. Peak disk is held to the temporary files in flight, one per worker: about 85 MB with a single worker and about 250 MB at the default four workers. The filtered Parquet output for apache/spark over that window is 53 KB, and local disk does not grow with the length of the window processed.
At full scale: a six-month window (January to June 2024, 4,368 hourly files, roughly 350 GB of compressed source) filtered to apache/spark completed in 11 h 16 min on the same laptop over a residential connection, with peak temporary disk of 565 MB and a 112 MB Parquet output. Peak in-flight disk runs higher than the six-hour benchmark suggests because hourly file sizes vary widely by time of day. The full reproduction, including a comparison against the study it reproduces, is at Spark_Codebase_Evolution.
- Python 3.10 or higher
pip
pip install gharcgit clone https://github.com/aravpanwar/gharc.git
cd gharc
python -m venv venv
# macOS / Linux:
source venv/bin/activate
# Windows PowerShell:
# .\venv\Scripts\Activate.ps1
pip install -e .gharc uses orjson instead of the standard library parser when it is installed. The fast extra pulls it in:
pip install "gharc[fast]"Download all activity for a specific repository over a one-day window.
Note that --end is exclusive, so this covers all 24 hours of 2024-01-01.
gharc download \
--start 2024-01-01 \
--end 2024-01-02 \
--repos "apache/spark" \
--output spark_data.parquet
For multi-hour or multi-day runs, prefer --output run.jsonl so the run can resume from where it left off if it crashes; convert to Parquet at the end with gharc convert run.jsonl run.parquet. See Resumable runs below for details.
Filter for multiple repositories and specific event types (e.g., only Pull Requests and Pushes). This covers all of June 2023 (June 1 inclusive through July 1 exclusive).
gharc download \
--start 2023-06-01 \
--end 2023-07-01 \
--repos "apache/spark, pandas-dev/pandas, pytorch/pytorch" \
--event-types "PullRequestEvent, PushEvent" \
--output oss_summer_2023.parquet \
--workers 4
| Argument | Description | Example |
|---|---|---|
--start |
Start date in UTC, inclusive (YYYY-MM-DD or YYYY-MM-DD-HH) | 2024-01-01 |
--end |
End date in UTC, exclusive (YYYY-MM-DD or YYYY-MM-DD-HH) | 2024-02-01 |
--repos |
Comma-separated repositories to keep; supports owner/* wildcards |
apache/spark,apache/* |
--orgs |
Comma-separated repository owners to keep | apache,pandas-dev |
--actors |
Comma-separated actor logins to keep | dongjoon-hyun,cloud-fan |
--event-types |
Comma-separated list of GHArchive event types | WatchEvent,ForkEvent |
--output |
Output filename (.parquet or .jsonl) |
data.parquet |
--workers |
Number of parallel download threads (default: 4) | 8 |
Dates are interpreted as UTC, matching GHArchive's hourly file naming.
Repository names are matched exactly and are case-sensitive, so pass the canonical owner/name as it appears on GitHub (for example apache/spark). Use apache/* or --orgs apache to keep every repository under an owner. --repos and --orgs are combined (an event is kept if it matches either), while --event-types and --actors further narrow the result.
For long jobs, gharc keeps a small <output>.state.json next to the output file listing which hours it has already processed. If the run crashes, restarting the same command picks up where it left off rather than redoing completed hours. The state file is written atomically (a temporary file renamed into place) so a crash mid-write cannot corrupt it, and it is removed automatically when the run finishes cleanly. Note that gharc relies on the operating system to flush writes to disk rather than forcing an fsync after each hour, so an abrupt power loss (as opposed to a process crash) could in rare cases leave an hour marked done whose data had not yet reached disk.
Resume support requires JSONL output. Parquet writers cannot append to a closed file, so for multi-hour runs use --output run.jsonl and convert to Parquet at the end:
gharc convert run.jsonl run.parquetThe CLI is a thin wrapper around gharc.process_range, which you can call directly:
from datetime import datetime
import gharc
gharc.setup_logging()
gharc.process_range(
start=datetime(2024, 1, 1),
end=datetime(2024, 1, 2),
repos=["apache/spark"],
event_types=None,
output="spark_one_day.jsonl",
workers=4,
)
gharc.jsonl_to_parquet("spark_one_day.jsonl", "spark_one_day.parquet")process_range also accepts orgs and actors keyword arguments mirroring the --orgs and --actors CLI flags, and entries in repos may use owner/* wildcards, so repos=["apache/*"] keeps every repository under an owner.
__all__ in gharc/__init__.py lists the public surface (process_range, jsonl_to_parquet, DataWriter, parse_date, date_range, get_url_for_time, setup_logging, plus the filter helpers).
For long date ranges, the included examples/orchestrator.py script runs gharc month by month so each year produces one Parquet file per month rather than one giant output:
python examples/orchestrator.py \
--start 2023-01-01 \
--end 2024-01-01 \
--repos "apache/spark,pandas-dev/pandas" \
--output-dir ./gharc_out \
--workers 4gharc/
├── src/gharc/ # Library + CLI entry point
├── tests/ # pytest test suite
├── benchmarks/ # Reproducible runs that back the performance claims
├── examples/ # Driver scripts (e.g. month-by-month orchestrator)
├── paper/ # paper.md, paper.bib, figures (the JOSS submission)
└── CITATION.cff # GitHub-detectable citation metadata
Contributions are welcome. Please read CONTRIBUTING.md for details on the process for submitting pull requests.
Running Tests:
pip install -e ".[test]"
pytest tests/The accompanying paper is at paper/paper.pdf and is rebuilt automatically on every push by the Paper CI workflow.
If you use gharc in your research, please cite it using the metadata in CITATION.cff or as follows:
@software{gharc2026,
author = {Panwar, Arav},
title = {gharc: A stream-and-filter tool for the GitHub Archive on consumer hardware},
year = {2026},
url = {https://github.com/aravpanwar/gharc}
}
This project is licensed under the MIT License - see the LICENSE file for details.
Created by Arav Panwar aravpanwar.com