README.md

metadata

license: cc0-1.0
pretty_name: Internet Archive census (April 2016) — per-file md5+sha1
tags:
  - file-hashes
configs:
  - config_name: data
    data_files: data/census-*.parquet
  - config_name: maps
    data_files: maps/md5.parquet

Internet Archive census (April 2016) — per-file md5+sha1

The April 2016 Internet Archive census recorded md5 and sha1 for every file of every public IA item — the only bulk file-hash inventory of the Archive ever published (the census was not repeated; nothing uploaded after 2016-04-11 appears here). This dataset is those two dumps joined — they are line-aligned outputs of one item walk — into single witness rows co-observing both digests, republished as page-indexed parquet for point lookups over HTTP range reads.

Every row's claim URL exists by construction: https://archive.org/download/<identifier>/<filename> (filenames are stored percent-encoded, exactly as the census dumps spell them — use them verbatim in URLs). Items can go dark after the fact, so a claim is "this item held these bytes in April 2016", and some URLs now 403.

Both digests are weak (md5, sha1): rows are corroboration and crosswalk evidence, never byte-identity proof. Downstream tools should render matches as "consistent with", not "is". hashdex does exactly that — its scan reports census matches in a separate "consistent (sha1)" tier.

150,129,712 rows across ~14.9M items; 129,730,596 distinct sha1s (equal to the md5→sha1 pair count: no sha1 in the corpus is claimed by two md5s). Items' own _meta.xml/_files.xml/.torrent sidecars (39.1M) are dropped as per-item plumbing; 496 rows with malformed hash fields skipped.

Layout

• data/census-{0..f}.parquet — sha1, md5, identifier, filename; sorted by sha1, sharded by its first nibble (~580 MB each).
• maps/md5.parquet — md5 → sha1, sorted by md5 (5.4 GB).
• ia-census.sha1.bloom — DCSO membership filter over the distinct sha1s (p = 10⁻⁴, 311 MiB, UPPERCASE-hex keys); hdx filters fetch ia-census installs it so scans can route matching files here.

All parquet carries page indexes (zstd, no dictionaries, 128 KiB pages): a point lookup over HTTP range reads fetches a few tens of KB.

Rebuild

ia_census_parquet.py in this repo streams the two census dumps (11.3 GB, from the IA item above), zip-joins them with per-line alignment checks, sorts one output file per sha1 nibble (DuckDB, memory-bounded), and transcodes with page indexes in per-file subprocesses (verified: sortedness, offset indexes, counts). The bloom is built by hdx from the distinct sha1 column:

hdx filters build ia-census sha1 sha1_keys.txt

The source census is frozen — a rebuild is deterministic.

Census data by the Internet Archive and the 2016 census effort, generously public; this compilation CC0.