feat: install a new server as a Docker installation (Linux)
Also give the installer's test helper the compatible_sha256 field that master
added to manifest::Migration after the pull request was branched.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
docker::install is the Linux counterpart of install::install_base and shares none
of its code (the Windows installer and its folder checks are untouched). Same
order of events: signed manifest, folder / space / game data checks, verified
download, extraction into <dest>.installing, database, one rename. Docker must be
usable; a failure removes the containers, the volume and the staging folder.
The databases start from the package's own dumps, Database/baseline/<kind>.sql.zst,
in the format of the Manager's backups, taken in a state where every migration of
the package is applied. The installer imports them, records the migrations as
applied (baseline) in a few bulk statements, and creates the acore account, the
realm port and the console account. Rebuilding from the repository's SQL does not
work: pending migrations such as rev_20260903_01_live_class_baseline are guards
generated from the maintainers' live database and refuse to run on a database
built from the base scripts, whatever the order.
- coa-release export-baseline --server DIR --out DIR produces the dumps from a
prepared server (repack or Docker).
- The game data stays where the person has it (dataDir in docker.json, mounted
read-only), checked at install time.
- ensure_main_configs creates worldserver.conf / authserver.conf from their
.dist templates when missing, at install and at every start, never overwriting.
- Fix (affects the backend merged in 0.6.0): the database container's health check
is now a ping over TCP. On a new volume the image first runs a temporary server
that answers on its socket but takes no network connection, so the container was
healthy about four seconds (10 s against 14 s measured) before TCP worked.
- docs/ARCHITECTURE.md: D13 (installation, package contract) and D14 (how Windows
and Linux code are kept apart).
Tested: 12 new tests (folder rules, settings, package checks, bulk ledger SQL,
failure cleanup with an in-memory docker), and real runs of the whole chain
through examples/docker_install: baseline dumps exported from a prepared
server, package signed with a throwaway key, installed in 171 s, started by the
Manager (database healthy, world ready in 8 s, auth up), console reached with the
account the installer created, 1630 migrations recorded as applied.
- config snapshots: two snapshots of the same scope in the same millisecond
shared one folder. A restore snapshots the current state just before reading
the snapshot it restores, so on a fast disk the snapshot being restored was
overwritten with the current files and the restore did nothing. Each snapshot
now gets its own folder (numeric suffix when the name is taken). This was the
intermittent failure of snapshot_restore_roundtrip_is_byte_exact_and_undoable.
- package extract: restore the execute bit on ELF binaries and #! scripts (the
archive carries no modes). No effect on Windows.
- fsx::starts_with_ci: case-insensitive on Windows only; /srv/CoA and /srv/coa
are different folders on Linux.
- registry test: use a root path that is a root on the running OS.
cargo test --workspace on Linux: 157 passed, 0 failed.
Verified on the fixture: signed update applied with a recovery point, template merge,
a real SQL migration, restart and readiness; failure/rollback paths covered by tests.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Verified: fixture packaged 2.2 GB -> 327 MB, installed in 18 s with rotated DB
credentials via the native UI; hostile archives, bad signatures and interrupted
downloads are rejected/resumed in tests.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>