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
Anyone could mint signed proxy URLs (a remote account changes its icon, an anonymous lookup returns the URL), and each
anonymous request held up to 40 MB in memory; the cache grew without bound between hourly trims; two clients asking
for the same new file downloaded it twice and wrote over each other in place, so a reader could get half a file with a
7-day cache header; a file over the limit was downloaded twice on every request; a failed fetch, a 404, was cached by
browsers for a week; cached media of a server suspended later were still served, and RejectMedia skipped avatars,
emoji, covers, video variants, link cards and remote edits; /clientapi/group/members returned remote pictures raw.
Now a download is shared by everyone asking at once, streamed into a .part file and renamed into place
(FederationHttp.DownloadMedia copies bounded, never into memory), at most eight at a time; a file too big to cache is
remembered for an hour and only streamed, a failure for five minutes; the cache's size is counted as it grows and
trimmed as soon as it passes the cap; a cached file is opened before it is answered; browsers may cache only a
success; nothing of a suspended server, or of one whose media are rejected, is proxied (everything remote a client
sees goes through the proxy, so that covers every kind), and blocking one purges its cache; the proxy has its own rate
limit per client address; group members' pictures are proxied; the proxy's key is loaded once, the oldest if two
were made. This changes what PrivaPub serves its clients, not what it sends to other servers.
Tests: clients asking at once share one download, a failure isn't cached by browsers, an over-limit file is fetched
three times for two requests instead of four, a blocked server's media are refused and its cache purged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
An upload went straight to libvips: whatever loader recognised the bytes ran (an SVG sent as image/png was rasterised),
nothing bounded how many pixels it would decode to (a small PNG could decode to gigabytes, three times over), a GIF
was loaded frame by frame and never resized, and all of it ran inside the request with nothing limiting how many at
once. A GIF was typed gifv but stayed a .gif, which a gifv player can't play; its metadata was kept; colours lost
their ICC profile without being converted; HEIC was advertised but the bundled libvips can't decode it.
Now:
- only libvips' JPEG, PNG, GIF, WebP and HEIF loaders ever run on an upload (every other loader is blocked);
- the header alone says how big an image would decode, refused above Media:MaxPixels (40 MP) or MaxFrames;
- a still image is shrunk on load, turned by its orientation and brought into sRGB (thumbnail), then written without
metadata, a profile picture the same way;
- an animated GIF becomes a looping silent H.264 mp4 typed gifv, as on Mastodon (PostMedia.Kind keeps it a gifv),
and a remote GIF is an image;
- processing runs Media:Concurrency at a time, and uploads have their own rate limit per credential;
- HEIC and HEIF are no longer offered.
Tests: only the upload formats load, the header tells the size, an SVG posing as a PNG and an image too large are
refused before decoding, an animated GIF becomes a gifv and a still one an image, HEIC isn't advertised. The media
scenarios against the pasture (GoToSocial, Mastodon, Misskey, Akkoma, Pixelfed, Smithereen, Vernissage, Castopod,
PeerTube) pass: 329 checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
A Follow was resent only when the persona followed again. Lemmy 1.0
sends nothing it queued for a server before it started sending there, so
the Accept of a community follow made on first contact was lost for good:
the village's persona stayed "requested" a day while Lemmy listed her as a
follower, and every post the community announced was refused as not
followed. A request still unanswered is now sent again, the same activity,
after 15 minutes, an hour, 6 hours, a day, two and four days
(FollowResender, every 15 minutes); a server that holds the follow answers
the copy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
PrivaPub now finds CDNs three ways, best first: the address ranges the
CDNs publish (Cloudflare, Fastly, Amazon CloudFront, Bunny, Gcore,
Imperva), downloaded daily by CdnUpdater and kept in CdnRangeSet; the
CDN's fingerprint in the responses it already gets from a server
(EdgeHintsHandler on the federation client); and the networks that carry
only a CDN. The fixed ASN list is gone; ASNs shared with plain hosting
(AWS, DataPacket) no longer hide a server. A server's Geo records the
CDN, its domain and how it was found, and weekly snapshots now keep the
city and coordinates too.
Servers through time (ServerPlaces): /instances/:host/history lists a
server's weekly snapshots, a CDN-fronted server's geo names the CDN's
domain and where the server was before it (before_cdn), and
/api/privapub/v1/cdns and /cdns/:domain group servers by CDN with week
by week who joined and who left. Owner decisions recorded in ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LsXgEaXee4GCU1hwYgPJXw
Owner decisions (2026-10-04, recorded in docs/ROADMAP.md): production runs everything that is built, and nothing waits
on a person running a command.
- Geolocation updates itself. GeoUpdater, a hosted service, checks daily whether each DB-IP Lite database was built this
month. If not, it fetches this month's, or last month's early in the month. It installs a file only once it opens as
the right kind of database, then swaps it in atomically, and the locator reloads at once. Lookups now run under the
lock, so a reload can no longer dispose a reader mid-lookup. The systemd timer, its script and their setup.sh lines
are gone: the root step they needed never happened, and none is needed now. /stargazing names the database in use.
- The admin CLI runs after the app is built, with every service and nothing started.
- `create-root <login> [--admin]` takes the password on stdin; it is how the first login is made while sign-up is
closed.
- `smoke <persona>` keeps the root `deploy-smoke` and an undiscoverable persona, and gives the root a new password
on every run.
- The deploy signs in as @thepra. It runs the CLI, gets a token through the real OAuth flow (tools/smoke/oauth.sh,
moved out of the pasture's privapub_token, which now uses it), checks the signed-in API and that @thepra is
undiscoverable, then revokes the token. PRIVAPUB_SMOKE_TOKEN is gone.
- The deploy also fails when:
- NodeInfo and the instance API disagree about registrations;
- /stargazing does not say the crawler is on;
- the geolocation databases are missing or more than 40 days old.
- The crawler is on in production, seeded with ten large servers of different kinds. FEDERATION.md now describes it
and how to opt out.
- One registrations switch (Registrations:Mode, default Invitations; Open in tests and the pasture). It is read by
open sign-up (403 when closed), NodeInfo `openRegistrations`, and v1 and v2 of the instance API, so they can no longer
disagree. Before, NodeInfo said open and the instance API said closed. Group invitations always work, so
invites_enabled is true.
- A persona edit through /clientapi no longer resets what the Mastodon API set (discoverable, locked, quote policy…):
the theme is merged into the settings instead of replacing them.
650 tests pass. The deploy's smoke step was rehearsed against the pasture's PrivaPub.
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
PrivaPubHost is a WebApplicationFactory<Program> on the fixture's database, configured only
through UseSetting (visible before Build, unlike ConfigureAppConfiguration). It drops the
background workers so tests run the jobs they queue (Jobs.Run, RunInbox), gives each client
its own address for the rate limiter, and has a SecureMode variant. Accounts signs up roots,
adds personas and gets Mastodon tokens through the real /oauth code flow. RemoteActor signs
HttpRequestMessages for the real /peasants routes; Peer records bodies and headers and
serves files with ranges and text pages.
Program registers the Guid serializer with TryRegisterSerializer, so a second host in one
process starts; the fixture runs the migrations in production's order before any test.
HostBootTests: the host shares the fixture database; the harness handles exactly the
activities the server registers; every job kind has one handler; every controller and page
model can be made; the service graph validates with ValidateOnBuild and ValidateScopes;
Swagger is 404 outside Development; a persona's token never names its root; a signed DM
through the real /mouth route is queued, processed and stored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ELjqpznMFMNrJoJUj6K5p2
thepraandClaude Opus 5.5