uv-scripts/data-processing

Data processing

Part of uv-scripts — self-contained UV scripts you run on Hugging Face Jobs in one command.

General data processing recipes: convert, clean and prepare data files.

Script	What it does
optimize-parquet.py	Converts CSV, JSON and Parquet files uploaded to a bucket into optimized Parquet, triggered by a bucket webhook

optimize-parquet.py: optimized Parquet from bucket uploads

Upload a CSV, JSON or Parquet file to a Storage Bucket and get an optimized Parquet version in a second bucket, automatically. A bucket webhook starts a Job for each upload, and the Job converts only the files that changed.

input bucket  ──upload──▶  webhook  ──▶  Job (optimize-parquet.py)  ──▶  output bucket
data.csv                                                                 data.csv/data/train-00000-of-00001.parquet

The output is written by datasets, so it gets the same optimizations as push_to_hub: content-defined chunking for Xet deduplication, a page index for fast filtering and random access, and row groups of at most 100 MB.

Setup

You need two buckets: one you upload to, and one for the output. The Job writes to a different bucket so that its own output does not trigger it again.

hf buckets create my-raw-files --private
hf buckets create my-parquet --private

1. Create a base Job for the webhook to re-run. With no webhook payload, this first run exits straight away:

hf jobs run --flavor cpu-upgrade --timeout 2h -e OUTPUT_BUCKET=<user>/my-parquet \
    ghcr.io/astral-sh/uv:python3.12-bookworm \
    uv run https://huggingface.co/datasets/uv-scripts/data-processing/raw/main/optimize-parquet.py

Use hf jobs run ... uv run <url> here, not hf jobs uv run <url>. hf jobs uv run uploads the script as a volume, and webhook runs don't keep volumes.

2. Create a webhook on the input bucket that re-runs this Job:

from huggingface_hub import create_webhook

create_webhook(
    job_id="<job id from step 1>",
    watched=[{"type": "bucket", "name": "<user>/my-raw-files"}],
    domains=["repo"],
    secret="<fine-grained token>",
)

The Job uses the webhook secret as its token to read and write the buckets. Use a fine-grained token, not your main one.

3. Upload a file:

hf buckets cp data.csv hf://buckets/<user>/my-raw-files/data.csv

After about a minute, the output is in <user>/my-parquet/data.csv/: the Parquet file(s) under data/, plus a README written by datasets.

Options

Environment variable	Default	Meaning
OUTPUT_BUCKET	required	Bucket to write the Parquet files to. Must differ from the watched bucket.
STREAM_ABOVE_BYTES	1/3 of free disk	Files larger than this are streamed instead of loaded to disk.

Files that fit on the Job's disk are loaded in full. Larger files are streamed, so they don't need to fit on the disk (50 GB on cpu-upgrade); raise --timeout for very large files. Supported inputs: .csv, .json, .jsonl, .parquet. Other files are skipped, and deleted files are ignored.

Notes

• Cost: a small file takes about 20 seconds on cpu-upgrade ($0.03/hour). In testing, one hf buckets sync of several files sent one webhook event, so it started one Job.
• Pin the script: each webhook run downloads the script again. To stop changes to this recipe from reaching your webhook, replace main in the URL with a commit hash.
• Limits: a webhook can trigger at most 1,000 times per 24 hours. Above 10,000 changed files in one event, the payload list is truncated; those files are not converted.