install-media / README.md
david-ar's picture
install-media dataset: 382,780 rows / 32 witnesses (2026-08-15 harvest)
ac7cc1e verified
|
Raw
History Blame Contribute Delete
3.61 kB
---
license: cc0-1.0
pretty_name: Install-media checksums (32 distros)
tags:
- file-hashes
configs:
- config_name: data
data_files:
- split: sha256
path: data/media-sha256.parquet
- split: sha512
path: data/media-sha512.parquet
- split: sha1
path: data/media-sha1.parquet
- split: md5
path: data/media-md5.parquet
---
# Install-media checksums (32 distros)
Per-image checksums of **OS install and live media as released**,
harvested from the distros' own checksum documents (SHA256SUMS and
friends) beside the images. Nobody aggregates these (quickget ships
per-distro endpoint *recipes*; osinfo-db has no hashes at all), so
this dataset does. Every row's claim URL is the checksum document
itself — the statement is the project's own.
**382,780 rows** across 32 witnesses, full release history wherever
the tree enumerates it, harvested 2026-08-15:
- **ubuntu** (releases + old-releases back to warty 4.10, MD5SUMS era
included) plus every cdimage flavor tree (kubuntu, xubuntu,
lubuntu, ubuntu-mate, ubuntustudio, ubuntu-budgie, ubuntukylin,
ubuntucinnamon, edubuntu, ubuntu-unity, ubuntu-gnome, mythbuntu)
- **debian** — current debian-cd/-live plus the whole cdimage archive
(~150 point releases)
- **fedora** — every release: dl (current), archives, Fedora Core
1–6 (SHA1SUM era), secondary arches
- **archlinux** (every monthly ISO via the releng JSON), **almalinux**,
**rockylinux** (vault), **centos-stream**, **linuxmint** (+ LMDE),
**kali**, **opensuse** (Leap history + Tumbleweed), **alpine**,
**openbsd**, **freebsd**, **netbsd** (per-release SHA512),
**void**, **qubes** (`.DIGESTS` co-stating md5+sha1+sha256+sha512 —
4-hash single-witness rows), **raspberrypi** (dated image trees),
**endeavouros** (sha512), **openwrt** (every release, all targets,
top-level images only), **proxmox**, **zorin**
Digest schemes are whatever each document states: modern rows are
sha256 (some sha512); pre-2007 media are md5/sha1-only and should be
treated as weak lookup keys, never merge keys.
## Layout
One table per key scheme, each holding every row that carries that
scheme, **sorted by it** (rows appear in several tables when one
document co-stated several digests):
- `data/media-sha256.parquet` — witness, sha256, sha512, sha1, md5,
name, version, filename, loc
- `data/media-sha512.parquet`, `data/media-sha1.parquet`,
`data/media-md5.parquet` — same columns, different sort key
- `install-media.sha256.bloom` — DCSO-format membership filter over
the 370,310 distinct sha256 keys (p = 10⁻⁴, lowercase-hex keys).
`hdx filters fetch install-media` installs it so `hdx scan` can
route matching files to this dataset.
- `media.tsv.gz` — the normalized harvest the parquet is built from
(one row per witness × scheme × file; the parquet regroups digests
only when one document co-stated them — sibling checksum documents
never merge)
`loc` is the URL of the checksum document stating the digest. All
tables carry page indexes (zstd, no dictionaries, 128 KiB pages), so
a point lookup over HTTP range reads fetches a few tens of KB. This
is how [hashdex](https://github.com/davidar/hashdex) answers
`hdx <iso-sha256>` with "Fedora releases these bytes as
Fedora-Workstation-Live-44-1.7.x86_64.iso" — from the CLI and from a
static browser page.
## Rebuild
`tools/install_media.py` (checksum-document GETs only, never
content) then `tools/media_parquet.py`, both in this repo and in the
hashdex GitHub repo. Trees are a point-in-time snapshot; re-harvest
refreshes them (delete the dumps cache first).