david-ar/rpm-files

RPM per-file digest index

Per-file content digests for every package in every RPM repository we could enumerate, harvested without downloading a single package: repo metadata (primary.xml) publishes each RPM's exact header byte range, one HTTP range GET per package fetches just the header, and its FILEDIGESTS tag states the digest of every installed file — aligned with the path arrays, so each row is "this repo ships these bytes as /usr/bin/bash in bash-….rpm". Package rows (the checksum of the .rpm artifact itself, primary.xml's pkgid) get their own table, so both an installed file and the rpm that shipped it resolve.

567,931,113 file rows across 3,772,141 packages / 348 witness repos, harvested 2026-08-15, enumerated live from the hosts' own directory listings:

• Fedora — Core 2–6 and Fedora 7–44, releases + updates, archives.fedoraproject.org included (Core-era rows are md5 — FILEDIGESTS predates sha256)
• EPEL — 4 through 10, archived point snapshots included
• CentOS — Stream 9/10 (BaseOS/AppStream/CRB) and the full vault: every point release 3.7 → 8.5
• AlmaLinux / Rocky Linux — current + full vaults (8.3 onward)
• openSUSE — Leap 15.0 → 16.x + update repos, Tumbleweed snapshot (sha512 pkgids)
• RPM Fusion — free + nonfree, current releases (no public archive of EOL release repos exists)
• Microsoft VS Code yum repo

This dataset supersedes the per-repo blooms in rpm-header-blooms.

Layout

Normalized, page-indexed (zstd, no dictionaries, 128 KiB pages), so point lookups over HTTP range reads fetch a few tens of KB:

• data/packages.parquet — pkgid, scheme, witness, name, evr, arch, location; one row per package × witness, sorted by pkgid (the rpm file's own checksum — a direct lookup key). location is the full URL of the rpm: a claim URL by construction.
• data/files-sha256-{0..f}.parquet — digest, path, pkg; sharded by first digest nibble, sorted by digest. pkg is the absolute row index into packages.parquet (the normalization: package identity stated once, referenced ~120× cheaper than repeating it).
• data/files-md5.parquet, data/files-sha1.parquet — same shape for the digest schemes ancient repos stated (weak lookup keys, never merge keys).
• rpm-files.sha256.bloom — DCSO-format membership filter over the 111,904,198 distinct sha256 keys, file digests and pkgids together (p = 10⁻⁴, lowercase-hex keys). hdx filters fetch rpm-files installs it so hdx scan routes matching files here.
• tsv/<witness>.tsv.gz — the raw harvest, one file per witness (witness, scheme, digest, name, evr, arch, path, pkgid, location; - = absent; package rows have path/pkgid -).

This is how hashdex turns a scan hit into "Fedora 44 ships these bytes as /usr/bin/bash in bash-5.3.9-3.fc44.x86_64.rpm" with the rpm's URL — from the CLI and from a static browser page.

Rebuild

tools/rpm_files_job.py (in this repo and in the hashdex GitHub repo) enumerates the repo families from their directory listings and runs the whole pipeline — harvest (repo-metadata GETs plus one header range GET per package, never content), normalized parquet build, bloom — as a Hugging Face Job (hf jobs uv run --flavor cpu-xl --timeout 10h -s HF_TOKEN tools/rpm_files_job.py). Per-witness TSVs upload as they complete and double as resume markers, so a relaunch continues where it stopped. Repos are a point-in-time snapshot; re-harvest = delete tsv/ and relaunch.