A backup is a directory <stamp>-<kind> under Backups:Root (/var/lib/privapub/backups, 2770, files 0640), written as
.partial and renamed once whole: a manifest (host, build, newest migration, each collection's count, size, sha256 and
indexes, what was left out and why, the media list), each collection as gzipped canonical Extended JSON read raw, and
hard links to the files of untrashed media rows (copies where a link can't be made). On a replica set every collection
is read in one snapshot session. Never in a backup: the statistics salt, jobs, recovery codes, sessions, the
maintenance lock and the configuration's copy with its SMTP password.
One backup or restore at a time (MaintenanceLock, a heartbeat document), and the janitor purges nothing meanwhile.
BackupScheduler backs up nightly at 03:30 UTC (or at once after missing a night); rotation keeps 7 daily, 4 weekly,
3 pre-deploy and 3 pre-restore backups. CLI: admin backup [--kind] [--db-only], admin backups, admin backup verify;
these run before migrations, so the deploy's own pre-deploy backup, which replaces mongodump, is of the database as the
live build left it. EntityMaps.Warm runs once under a lock, since test hosts now boot side by side.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
tools/pasture/flood/flood.cs answers as twenty fake servers (flood1..20.test)
and sends signed Creates, Likes and Follows at a set rate; load.sh measures
the answers, the queue's wait and processing times, its drain and a persona's
home timeline meanwhile (docs/LOAD.md has the method and the runs).
What the runs found:
- Every unique index was partial on $type: "string", which MongoDB never uses
for an equality lookup, so every post by ObjectURI, actor by ActorURI,
deleted object, domain block, remote instance and the rest was a
collection scan (280 ms a post lookup at 30 000 posts). They are partial on
$gt: "" now, which an equality on a string implies; MongoDB.Entities
rebuilds them in place at the next start.
- Two inbox workers capped intake near 110 activities a second:
Federation:InboxConcurrency and DeliveryConcurrency (default 8) set them.
- The indexes the plan listed as missing: a post's boosts and replies, a
persona's boosts, who follows an actor, timeline rows by author, a post's
likes and pins.
At 300 activities a second (200 let through, the rest 429 by the per-origin
limit) the queue wait went from 29 s to 6 ms at p50; with the limits lifted
PrivaPub processes about 900 a second, each in under 10 ms, and the home
timeline stays under 20 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
tools/pasture/town/ (run through tools/pasture/town.sh) seeds a fake
community across every running peer and checks that all of them, and
PrivaPub, agree on what happened:
- drivers per platform on four dialect bases (Mastodon API, Misskey API,
Lemmy API, PrivaPub with /clientapi), each with a selftest against
its own server; what a server holds is read from its database, never
by making it fetch;
- a deterministic generator (specs/village.json: 23 accounts on seven
servers, roots with several personas, circles and communities, a
cross-server follow graph, posts of every kind and visibility, reply
rounds, likes, boosts, reactions, votes, edits, deletes, blocks,
mutes and a report) and a seeder that keeps a ledger of what happened;
- a sweep that expects delivery and confinement per server, what each
account sees, counts, threads, edits, deletes, follows and privacy
rows (sibling keys, published days, canary root credentials in every
peer's database, located posts that never leave), with what the peers
do on purpose modelled (Misskey drops orphan replies, Lemmy keeps only
community content, edits go to the post's own audience);
- known gaps (gaps.json) turn failures into xfail and passes into xpass;
a self-contained report.html, and docs/INTEROP-BACKLOG.md.
The pasture moves to a public-looking subnet (peers with no private
address switch can join), takes PASTURE_PORT when 6971 is in use, adds
peers to a running pasture (run.sh add, Caddy recreated with its CA
kept), removes its volumes on down, writes every scenario check to
out/scenarios.jsonl, serves decePub as decepub.test for its e2e tests,
lifts GoToSocial's and Lemmy's own rate limits, trusts Caddy in
Mastodon (TRUSTED_PROXY_IP) and gains Hollo (Fedify), whose one login
owning several accounts is the nearest peer to PrivaPub's personas.
The first village found the four PrivaPub bugs fixed in the commits
before this one; the second run, on the fixed server, passes 2319 checks
with 11 failures left, all between peers or from Lemmy's send worker,
which the seeder now warms up first. ROADMAP records the owner's
decisions of 2026-10-04 (the town, and P9 back from the cut list).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
tools/pasture/run.sh up [peer...] | down | logs | ps and interop.sh [peer...] are now built
from parts:
- lib/pasture.sh: network, Caddy with its CA in a volume and copied to .ca/root.crt, Mongo,
PrivaPub;
- peers/<name>.sh: each peer's <name>_up;
- lib/interop.sh: ok, ko, and xf for checks expected to fail until a later phase;
privapub_token <persona>, which gives each peer its own persona under one root; and
stats_check;
- scenarios/<name>.sh: each peer's checks.
GoToSocial is pinned at 0.22.1, the version the 33 checks were proven on. Its scenario
gains four statistics checks:
- the admin statistics describe gts.test as gotosocial;
- they count inbound and outbound traffic;
- they name no account.
37/37 passed three runs in a row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ELjqpznMFMNrJoJUj6K5p2
A RollupDay job folds each finished day's InteractionEvents into one InstanceDay per server:
- Counters for the admin, keyed channel:activity:object:outcome:reason, plus signature
schemes, audiences, local kinds and features;
- PublicCounters, everything a public page may ever read:
- inbound and outbound activities from an allowlist, on public, unlisted or unaddressed
traffic only, with the outcome collapsed (accepted or dropped, delivered or failed);
- no reasons, no Flag or Block;
- features of public objects;
- health:ok or health:failed from reachability (a 4xx means the server answered);
- latency, wait and byte histograms, and the number of distinct accounts from that day's
hashes.
Re-running a day replaces it and keeps the live Reads counters. The day's salt is then
deleted, so its hashes can never be recomputed, and the next day is queued.
StatisticsSchedule plans today's rollup every hour and catches up any of the last seven
days that have events but no rollup. Waits round up into their bucket.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ELjqpznMFMNrJoJUj6K5p2
tools/pasture/run.sh starts PrivaPub, GoToSocial and Mongo on one podman network behind Caddy's internal CA, and
interop.sh drives both through their own client APIs: follows (one to a locked account), posts, CW, replies, likes,
boosts, DMs, edits, deletes and unfollow. All 25 checks pass, three fresh runs in a row.
- Federation:AcceptAnyCertificate joins the two test-network switches; startup refuses all three in Production.
- WebFinger falls back to http only when AllowPlainHttp is on.
- A bootstrap logger, so a failure before the host is built is no longer silent.
- P4 is ticked in the roadmap, with what has not been run live (Lemmy, a Mastodon circle member).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012CzABvBkbcFqoHdmi8b9WB
- /api/v1/media and /api/v2/media (and GET/PUT /api/v1/media/:id):
images go through libvips (NetVips, its native build bundled):
autorotated, every kind of metadata dropped (EXIF, GPS, XMP, IPTC,
comments), capped at 4096 px, with a 640 px preview and a blurhash
(own encoder, the reference algorithm); animated GIFs are re-encoded;
video and audio are remuxed by ffmpeg with -map_metadata -1, never
re-encoded, and a video gets a still preview. Files get random names
under /var/lib/privapub/media, outside the web root deploys replace, and
are served at /media/files with nosniff and a sandbox CSP.
- media_ids on create and edit (four at most, the persona's own, each used
once); notes carry them as Document attachments with alt text, blurhash,
focalPoint and size; inbound attachments were already kept.
- avatar and header uploads in update_credentials, cropped to 400x400 and
1500x500, federated with Update{Person}.
- Remote media reaches clients only through /media/proxy/{hmac}/{url},
fetched by the guarded client (no SVG, 40 MB cap) and cached outside the
served root, trimmed to 5 GB; foreign avatars and headers use it too, so
a client never contacts another server.
- MediaJanitor deletes uploads left unattached for a day.
- nginx accepts 100 MB bodies on the upload endpoints only (applied on Max).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012CzABvBkbcFqoHdmi8b9WB
thepraandClaude Opus 5.5