Fixed
- The shipped image can now actually drive the control plane: the release Dockerfile installs
docker-ce-cli+docker-compose-pluginfrom Docker's apt repo (Debian'sdocker.ioships CLI 20.10 without compose v2, so every service start/stop/logs/update call failed with'compose' is not a docker command— silently). - Region apply works end-to-end. The Phoenix app now ports the Go sidecar's full apply pipeline: PBF download (streaming, skip-if-present,
.partial+ rename) → GTFS download (non-fatal) →current.osm.pbf(symlink for one source, nativeosmium mergefor several) →current.osm.bz2for overpass → OTP staging (region.osm.pbf+ GTFS zips,graph.objdropped) →docker compose restartof the enabled ingest services. Previously no component downloaded PBFs at all and osmium ran viadocker run -v /data:/dataagainst a host path that doesn't exist. - Control-plane errors are no longer swallowed.
docker composeexit codes propagate intoservices.last_errorand an:errorstatus (a failed stop no longer pretends the service is "stopped"); region-apply failures broadcastapply_errorand persist inRegionApplier.status/0, so the map page shows the real failure instead of "Applying N regions…" forever. - Region selection no longer accumulates invisibly: the Settings Region tab shows a removable-chips tray of every selected region with "clear all"; the apply button counts only actual changes vs. the last applied selection, and the flash names the regions it applies.
- Tile-pack downloads are asynchronous (no 30-minute
GenServer.callceiling), report real byte progress fromContent-Length, land indata/tiles/where Caddy serves them (/tiles/*— previously they were written todata/app/tiles, which nothing served), and require a size confirmation (HEAD probe) before multi-GB fetches. - Log streaming works in both viewers: the map-page logs modal now streams real lines (was: a single
last_logline) with waiting/EOF/error states; the admin viewer starts with 200 lines of history (--tail=200, was--tail=0) and announces stream end instead of freezing. Log tailers register uniquely (no duplicatedocker compose logsprocesses per open viewer) and no longer restart in a loop when the CLI is broken. - Settings panel boot race: a not-yet-ready control plane renders a "starting…" placeholder instead of a fake-empty "region: none" with a disabled button.
- Cosmetics: never-started services read "off" instead of "unknown"; the header region stat uses catalog labels ("Berlin +2") instead of raw
gf:-prefixed slugs; admin error states render trimmed human messages instead ofinspect/1terms. - The service logs modal overlays the whole page and actually closes. It was confined to the side panel (
absoluteinside the panel container) and anonclick="stopPropagation()"handler swallowed every click — including the close button — before LiveView's delegated listener saw them. The modal now renders at the page root (fixed, full viewport) and closes via the ✕ button, clicking outside, or Escape. - "Save & apply" no longer freezes (or appears dead) while docker pulls an image. Service enable/disable ran
docker compose up/stopsynchronously through GenServer calls (a 5-second LiveView call timeout against a multi-minute image pull). The compose op now runs in a background task: the click round-trips in milliseconds with an optimistic status, and the result (including failures) lands asynchronously inlast_error/status. - Enabled services come back after
docker compose downor a redeploy. At boot every service reconciles desired state (the persistedenabledflag) against the actual containers: enabled-but-gone services are started again, stale statuses ("ready" with no container, "stopped" with a live one) are corrected. - Service status updates without anyone watching logs. The status/progress parser was only fed while a logs viewer was open; a log tailer now attaches automatically to every running service. Tailers keep a 500-line ring buffer that's replayed into the logs modal and admin viewer when opened late — no more empty "Waiting for log output…" on a quiet, healthy service.
- Long-running operations are now visible in
docker logs atlas-app: region applies log start/phase/finish/failure, tile-pack downloads log start, one line per percent, and the outcome — previously a 100 GB planet download produced zero log evidence.
Added
- Raw Photon passthrough API under
/api/v1/photon/{api,reverse,lookup,status}: forwards the query string verbatim (repeated keys likeosm_tagpreserved) to the internal Photon service and returns its status and body untouched — no normalization, no{data, meta}envelope. Lets external proxies (e.g. chibigeo) offer a byte-faithful Photon-compatible API on top of Atlas. Photon errors pass through verbatim; an unreachable Photon yields503 UPSTREAM_UNAVAILABLE. - Control-plane preflight diagnostics (
Atlas.Control.Preflight): docker CLI, compose plugin, socket access, data-dir writability, and osmium are probed at boot; failures render a "Control plane degraded" banner in Settings and/admin/serviceswith the exact remedy (including per-OSDOCKER_GIDguidance — macOS OrbStack/Docker Desktop needsDOCKER_GID=0). - Live region-apply progress on the map page and
/admin/apply: per-phase card (downloading with byte %, merging, converting, staging, restarting) on a stablecontrol:applyPubSub topic; survives page refresh viaRegionApplier.status/0. GTFS_URL/GTFS_NAMEare now part of the region catalog (parsed from region.envpresets) so transit feeds download during apply.- Writable mounts for
data/gtfs,data/otp,data/tilesin both compose files. - App version display —
Atlas.Versionreads the canonicalmix.exsversion from the application spec (no separate version file to keep in sync) plus the git SHA baked in at image build time (APP_REVISIONbuild arg, passed by CI). Shown in the side-panel footer and the admin sidebar (both link to GitHub releases) and served atGET /api/v1/versionas{"data": {"version": ..., "revision": ...}}. - Dokploy deployment (
compose.dokploy.yml+DEPLOY-DOKPLOY.md) — Dokploy-managed services (no docker socket; the app reaches upstreams via env URLs), Caddy behind Traefik for TLS/domain, named volumes persisting across redeploys. Ships minimal —app+caddypointed at an existing Photon viaPHOTON_URL— with routing/POIs/transit (valhalla/overpass/otp) as uncomment-to-add blocks that build from OSM data staged by the map's region apply. Self-hosters keep the full click-to-enable control panel viacompose.yml. - OneDev CI build spec (
.onedev-buildspec.yml) — builds the app image with buildx on pushes tomain(APP_REVISIONbaked in viamoreOptions) and pusheslatest+ commit-SHA tags to OneDev's built-in registry atonedev.dwri.xyz/atlas/app; requires anonedev-access-tokenjob secret. Used for testing; the GHCR/production pipeline stays in.github/workflows/.
Changed
HOST_PROJECT_DIRis now consumed asdocker compose --project-directory(it was previously set but read by nothing), so sidecar bind mounts (./data/photon, …) resolve against the host checkout. Region-data processing (download/merge/convert/staging) no longer needs host-path translation at all — osmium runs natively in the app container, replacing the amd64-onlystefda/osmium-tooldocker-run dependency.