Revert VPS FTP after it crash-looped the app; document the diagnosis

The ltv account fails to authenticate from the VPS specifically
(works fine from Rocky with the same credentials, network path to
ftp.meteo.lv:21 confirmed fine from the VPS too) — narrowed to either
a password transcription error in .env.staging or LVGMC IP-allowlisting
the account, neither confirmable without inspecting the password.
Caught via the overnight watch this was set up for rather than
assumed safe; reverted ENABLE_LVGMC_FTP_JOBS to false on the VPS
immediately once the failure was confirmed real, restoring stable
operation. FTP stays enabled on Rocky, where it works.
This commit is contained in:
b0txec
2026-08-23 22:14:58 +03:00
parent 9a6cb30552
commit 1086c8b548
3 changed files with 49 additions and 13 deletions
+3 -2
View File
@@ -6,7 +6,7 @@ This directory contains the working documentation for the WeatherTool modernizat
- Windows is restricted to source editing, review, and Git operations. Rocky is the sole compile, build, development-runtime, and test environment. The Ubuntu VPS is a deployment target only.
- A production-like staging copy runs through Docker Compose on Rocky Linux at `http://192.168.1.101:9190`.
- Both Rocky and the VPS now ingest real data from two independent LVĢMC sources at once: the free `data.gov.lv` open-data feed and the private FTP feed (`ftp.meteo.lv`, real credentials obtained and verified 2026-08-23). The two are gated independently (`ENABLE_LVGMC_FTP_JOBS`, `ENABLE_HARMONIE_JOBS`) since real credentials for FTP and DMI HARMONIE arrive on different timelines — FTP is on; HARMONIE's credentials turned out not to be needed at all (DMI dropped its API key requirement — confirmed live, see the roadmap changelog) but wiring it up is deferred to a dedicated verification session, since the GRIB-parsing and map-rendering code has real, untested risk (a hardcoded crop/rotation calibration that may not match the current model grid).
- Rocky now ingests real data from two independent LVĢMC sources at once: the free `data.gov.lv` open-data feed and the private FTP feed (`ftp.meteo.lv`, real credentials obtained and verified 2026-08-23). The VPS currently runs open-data only — FTP was briefly enabled there too but reverted the same night after the `ltv` account failed to authenticate from the VPS specifically (see below). The two providers are gated independently (`ENABLE_LVGMC_FTP_JOBS`, `ENABLE_HARMONIE_JOBS`) since real credentials for FTP and DMI HARMONIE arrive on different timelines. HARMONIE's credentials turned out not to be needed at all (DMI dropped its API key requirement — confirmed live, see the roadmap changelog) but wiring it up is deferred to a dedicated verification session, since the GRIB-parsing and map-rendering code has real, untested risk (a hardcoded crop/rotation calibration that may not match the current model grid).
- A real UTC-vs-local timezone mismatch between the two sources (open-data's timestamps were UTC, FTP's already local, both stored in the same column with no conversion) was found and fixed before enabling both together; both `weather` tables were backed up and wiped for a clean, consistently-timestamped restart.
- The safe scheduled jobs (open-data station ingestion, GRIB cleanup) always run; the FTP job now runs too since real credentials exist; HARMONIE's job stays off pending its own implementation work.
- PostgreSQL is private to the project Compose network; only the Scala application publishes a host port.
@@ -21,7 +21,8 @@ This directory contains the working documentation for the WeatherTool modernizat
- The VPS `weather` table was wiped twice on 2026-08-23: first to remove the original 14-day synthetic dataset (backed up to `/srv/weathertool/backups/pre-real-data-release/`), then again after the UTC/local timezone fix (backed up to `/srv/weathertool/backups/pre-timezone-fix-wipe/`) so the two real sources it now holds — open-data (minutes 15/45) and FTP (minutes 11/13/23/30) — are consistently timestamped from a clean start.
- Approved 1920×1080 and 3840×1440 PNG production bases are now the rendering source for Faktiskā and Ūdens temperatūra; code draws only the changing values, selected weather symbols, and wind data over those fixed newsroom graphics.
- Browser branding assets and Latvian Open Graph/Twitter metadata are included for favicon, Apple home-screen icon, and link-preview support. Public crawler access still depends on the Nginx/Authelia policy used for the metadata and preview image.
- Browser and API verification is complete for the deployed `3eddf95` release: exact release image smoke-tested on Rocky before transfer (bundle hash and correct-local-time API responses matched the known-good local build), checksum verified on both ends, container health/loopback/public HTTPS confirmed. VPS's first scheduled FTP fetch specifically (not just open-data, which fires more often) is being watched for overnight to confirm outbound FTP actually works from the VPS network before relying on it unattended. Newsroom testers will be given the domain first, without walkthrough, to see how intuitive the workflow is before any guidance is provided.
- Browser and API verification is complete for the deployed `3eddf95` release: exact release image smoke-tested on Rocky before transfer (bundle hash and correct-local-time API responses matched the known-good local build), checksum verified on both ends, container health/loopback/public HTTPS confirmed.
- **VPS FTP is currently disabled again** (`ENABLE_LVGMC_FTP_JOBS=false`), despite being enabled earlier tonight. Watching for its first scheduled fetch (rather than assuming it was safe just because Rocky worked) caught a real problem: the `ltv` account fails to authenticate from the VPS specifically, and because a scheduled-task failure cascades through `parMapN` up to the app's top-level error handler, this was actively crash-looping the entire app in-process (not a Docker-level restart, but the whole app — DB, HTTP server, schedulers — tearing down and rebuilding every ~5s, repeating every ~2 minutes as the schedule re-fired). Reverted immediately; confirmed stable afterward. VPS network connectivity to `ftp.meteo.lv:21` itself is fine (verified with a raw connection test); this is specifically the `ltv` account failing from the VPS's IP, not a firewall issue — see the roadmap changelog for the full diagnosis and next steps (check the password transcription, or ask LVGMC whether the account is IP-restricted). FTP stays enabled on Rocky, where it's confirmed working.
- Frontend dependency maintenance is complete: Solid runtime and Vite tooling were updated, obsolete packages were removed, TypeScript checking was added, and a clean Rocky `npm ci`, typecheck, production build, full audit, and production-only audit all pass with zero known vulnerabilities.
- This is not yet approved or hardened for workplace production.