Remove deploy/nginx and the nginx service from compose.prod.yml: the
shared edge in Infrastructure owns ports 80/443 and serves every
domain. Make the deploy job fail when zitadel-init or seaweedfs-init
fail, and document the runner and edge setup.
Postgres, Zitadel, SeaweedFS, RabbitMQ and Jaeger no longer run as part
of this stack. They're shared services (used by Musify, Ping and
docker-manager) that now live in a separate Infrastructure repository
and must already be running.
deploy/compose.yml keeps only what's specific to Musify: its own
database connection, its own OIDC app and storage bucket (created
against the shared Zitadel/SeaweedFS via zitadel-init/seaweedfs-init),
and the application containers. A new wait-for-postgres one-shot
replaces the in-file Postgres health check, which no longer works now
that Postgres lives in a different compose project.
compose.prod.yml joins Infrastructure's infra-net network instead of
running its own copies. Docs and READMEs updated to match.
Drops the YouTube-backed second catalog source entirely: the
Local/External split on Track/Album collapses into single concrete
entities, Artist/TrackArtist/ExternalTrack/ExternalAlbum and the
YouTube service/download pipeline (yt-dlp, MassTransit consumer,
routing slip) are deleted, and mixes now source discovery only from
local listening history. Includes the RemoveYouTubeIntegration EF
Core migration and matching Worker/deploy/compose cleanup (no more
yt-dlp/python/deno in the Worker image, no thumbnail-cache/YtDlp
config sections).
Removes every wrapper script under deploy/ (up.ps1, up.sh, scripts/*,
zitadel/provision.*) and folds their logic into three one-shot services
in compose.yml, following the pattern already used by seaweedfs-init:
keys-init generates the RS256 stream-ticket key pair
migrate applies the EF Core migrations (backend/Musify.Migrator.Dockerfile)
zitadel-init creates the Musify project + OIDC apps, writes the client
ids back into the active env file
Deploying is now two plain `docker compose` commands, no scripts to run
or chmod:
docker compose -f deploy/compose.yml -f deploy/compose.dev.yml up -d
docker compose -f deploy/compose.yml -f deploy/compose.dev.yml --profile apps up -d --build
(swap compose.dev.yml for compose.prod.yml + --env-file deploy/.env.prod
for production). The first command brings up infra and lets the bootstrap
jobs run to completion; the second builds and starts the apps, which by
then read real OIDC values instead of empty ones — compose interpolates
${VAR} once per invocation, so this can't collapse into a single command
without the apps reading their own config file at container start instead
of via compose env (out of scope here).
Also:
- DatabaseDesignTimeFactory now also reads Database__ConnectionString
from the environment, and DesignSettings.json is optional, so the new
migrate container works without user-secrets or a config file excluded
from the Docker build context by .dockerignore.
- nginx's depends_on no longer lists the apps-profile services: it does
not need them at its own startup (every vhost resolves its upstream at
request time via the embedded Docker resolver), and requiring them made
`compose.prod.yml up -d` (without --profile apps) an invalid project.
- env_get/env_set in zitadel-init strip \r: the mounted env file can have
CRLF line endings on Windows, which otherwise makes an empty-looking
value non-empty to `[ -z ... ]`. .env.example/.env.prod.example are
pinned to LF via .gitattributes for the same reason. Also fixed:
rewriting the bind-mounted env file in place instead of renaming a temp
file over it (rename onto a bind-mounted inode fails with "Resource
busy").
- Removed compose.share.yml and zitadel/.output-share/ — a "temporary"
personal overlay that had leaked a real Zitadel service-account PAT
into git history (committed in 70968b3). That token should be rotated.
- Removed deploy/.tmp/ from .gitignore: nothing mounts or creates it
anymore (it never actually did, either — no compose service referenced
it, only the now-removed up.ps1 did).
Verified end-to-end with Docker: a clean `down -v` + fresh `up -d` +
`--profile apps up -d --build` on the dev stack brings up all four one-shot
jobs successfully and all four apps healthy (api/gateway return 200,
worker starts its consumers with no errors, web-player responds). Also
verified compose.prod.yml starts nginx cleanly with dummy certs and no
apps running, confirming the depends_on fix.
Thumbnail responses are immutable for a month, so nginx can answer repeats
from disk and never wake the web player or Google for them. The cache lives in
its own volume, collapses concurrent misses with a lock and serves stale
copies while it refreshes; THUMBNAIL_CACHE_SIZE sizes it.
The dev and prod stacks each declared the same nine services, which is how the
S3 credentials and the gateway configuration ended up diverging. compose.yml
now holds every service with no published port and no domain in it, and the
overlays only add what is specific: host ports and pgAdmin for development,
the nginx edge and the mounted secrets for production. The applications sit
behind the `apps` profile, so a plain `up` still starts infrastructure only.
Everything environment-specific moves to the env files: the public URLs, the
credentials and the environment name. The SeaweedFS identity file is rendered
inside the container from S3_ACCESS_KEY and S3_SECRET_KEY, so the store and
the applications can no longer be given different keys.
Admin