RST-4402: restore of /srv/atlas came back larger
Requested 2026-05-21 · Tape LTO-6, barcode A0071L6, written
2019-06-17 · Requested by platform, for a licence audit
Why we restored it
Legal needed the /srv/atlas tree as it stood in June 2019 to answer a
question about which build of a library was in the image. Routine. The tape is
one of forty-one in the 2019 offsite set, all written by the same appliance, all
verified on write and re-verified in the 2022 audit.
What came back
| On the manifest | Restored | |
|---|---|---|
| Files | 118,406 | 119,847 |
| Directories | 3,912 | 3,913 |
| Bytes | 44,109,882,113 | 44,116,551,204 |
| Checksum (manifest) | c41f…8ae0 |
c41f…8ae0 |
The manifest checksum matched. The manifest checksum is computed over the file list, and the file list is on the tape, and the file list has 118,406 entries. The restore wrote 119,847 files. Both numbers are correct; they disagree.
The extra directory is /srv/atlas/held/. It contains 1,441 files, which is the
difference. Nothing on the manifest mentions it. Nothing in the 2019 source
control mentions it. atlas was decommissioned in 2021 and its owners have all
left.
What is in it
Every file in held/ is named for a date and has the same shape: a plain-text
recursive listing of /srv/atlas, with sizes and mtimes, as it stood on that
date.
$ ls held/ | head -5
2015-08-30.list
2015-08-31.list
2015-09-01.list
2015-09-02.list
2015-09-03.list
$ ls held/ | tail -3
2019-06-15.list
2019-06-16.list
2019-06-20.list
$ ls held/ | wc -l
14411,441 dates, consecutive from 2015-08-30 to 2019-06-16, and then
2019-06-20.list. The tape was written on 2019-06-17.
The listings are accurate. We spot-checked eleven of them against the four independent backups we still hold from 2016, 2017 and 2018, and every path, size and mtime matched.
Every listing includes held/ itself, with a file count for that date. The
counts increase by one per day, exactly, from 1 on 2015-08-30.
2019-06-20.list reports held/ as containing 1,441 files.
Disposition
The tree has been copied to /srv/audit/rst-4402/ and the tape returned to the
vault. Legal got their answer from lib/, which is what they asked for, and the
ticket can be closed on that basis.
I have not opened 2019-06-20.list beyond held/. I am aware that this is not
a technical decision.
Related: stories/mtime, stories/inventory-diff. On why a stored artefact and a claim about it are two different objects, machinery/anatomy-of-a-page.