User reported: "배포가 되면 인증이 풀린다" (auth resets after every
deployment).
Root cause: Program.cs had no explicit Data Protection configuration.
Without SetApplicationName, ASP.NET Core derives the key-ring
discriminator from the app's physical content root path. Every
deployment lands in a brand-new directory
(~/deployments/quantengine_{tag}_{hash}/), so the discriminator
changed on every single release. The cookie authentication ticket is
encrypted/signed via this key ring, so once the discriminator
changed, every previously-issued auth cookie became undecryptable --
forcing all logged-in users to authenticate again after each deploy,
even well inside their 12-hour ExpireTimeSpan.
Fix: explicit .SetApplicationName("QuantEngine") pins a stable
discriminator across deployments, and .PersistKeysToFileSystem points
at %LOCALAPPDATA%/quantengine-keys (Linux: ~/.local/share/quantengine-keys
via User=kjh2064 in the systemd unit) -- a location outside the
versioned deployment directories, so the actual key material also
survives every redeploy and service restart instead of only the
discriminator being stable.
Discovered while verifying the new Operations page against a local
instance (SSH-tunneled to prod Postgres): every startup logged
'Hangfire setup failed: Cannot resolve scoped service
QuantEngine.Web.Services.SchedulerService from root provider' and
silently skipped InitializeSchedules() entirely.
SchedulerService is registered AddScoped, but UseHangfireSetup()
resolved it directly from app.Services (the root/singleton-level
provider), which cannot construct scoped services without an active
scope. This has apparently been broken for a while -- the 4 recurring
jobs (daily-collection, hourly-price-update, weekly-report,
monthly-optimization) only kept showing up because Hangfire persists
recurring job definitions in PostgreSQL from whatever earlier
deployment last managed to register them; any code change to those
schedules would silently never take effect on redeploy.
Fixed by creating an explicit scope (serviceProvider.CreateScope())
before resolving SchedulerService. Verified locally: the warning is
gone and the log now shows "Hangfire schedules initialized
successfully" followed by the dispatchers starting.
Root cause: user asked why logout was missing. Playwright audit against
production found logout works fine, but surfaced two real defects and
led to a wider audit that found extensive fabricated data across the
admin pages -- none of it backed by the database despite CLAUDE.md's
policy that all data must come from DB records.
Layout (_AdminLayout.cshtml):
- Full rewrite using Tabler's actual navbar-vertical/page-wrapper/footer
component structure instead of ad-hoc inline CSS. The old layout had
no footer element at all, and its mobile breakpoint CSS hid the
sidebar off-screen (left: -260px) with no hamburger button to bring
it back -- verified via Playwright screenshot at 375px width that
the entire nav menu was inaccessible on mobile, leaving only Logout
reachable. Tabler's navbar-toggler + Bootstrap collapse (bundled in
tabler.min.js) now restores it; verified the toggle actually opens
the menu via Playwright.
- Active nav-link highlighting moved from client-side JS string
matching to a server-side Razor helper against Context.Request.Path.
Fake/hardcoded data removed or replaced with real DB/Hangfire state:
- Dashboard: deleted the "최근 시스템 이벤트" table (3 rows hardcoded
from DateTime.Now with fake descriptions like "시스템 초기화" /
"데이터베이스 백업" -- no backing table exists). Removed hardcoded
"정상"/"연결됨" status badges and "버전: v0.1.0"/"업타임: 정상";
replaced with a real IsDatabaseConnected flag (true only if the
page's actual DB queries succeeded) and the real
IWebHostEnvironment.EnvironmentName.
- Monitoring: removed hardcoded "API 서버: 운영 중" (no real signal
backs it) and wired "데이터베이스: 연결 정상/끊김" to the same
real success/failure state as the page's own DB calls.
- Operations: this page was entirely fabricated -- ScheduledJobs,
RecentExecutions, IsJobProcessorRunning, PendingJobsCount, and
StatusMessage were all static values with zero connection to
Hangfire, despite Hangfire actually running in production
(confirmed via journalctl: ServerWatchdog, RecurringJobScheduler
dispatchers active) with 4 real recurring jobs registered in
SchedulerService (daily-collection, hourly-price-update,
weekly-report, monthly-optimization). Rewrote to query
JobStorage.Current.GetConnection().GetRecurringJobs() and
GetMonitoringApi() directly: real scheduled jobs, real succeeded/
failed executions, real server count, real enqueued count. Verified
locally (SSH-tunneled to prod DB) that this now returns the actual
4 registered jobs with correct next-run times and one real
RunDailyCollectionAsync execution.
Also fixed the page-title duplication on Monitoring/Operations
(ViewData["Title"] included "- QuantEngine" AND the layout appended
it again -> "모니터링 - QuantEngine - QuantEngine" in the browser tab).
Separately discovered (not fixed in this commit, flagging for
follow-up): Hangfire's SchedulerService.InitializeSchedules() fails
every startup with "Cannot resolve scoped service 'SchedulerService'
from root provider" -- the 4 recurring jobs above still show up
because they persist from an earlier successful registration, but
re-registration is silently broken on every current boot.
Verified end-to-end with Playwright against a local instance (SSH
tunnel to production Postgres): login, all 5 admin pages render
without errors, mobile hamburger opens the sidebar, and Operations
shows genuine Hangfire data.
User caught this directly: it's already 2026-07-12 in Korea, but
Run #2008's release was tagged quant_20260711.3.7150737 -- the wrong
date.
Confirmed: UTC was still 2026-07-11 16:2x when KST was already
2026-07-12 01:2x (9-hour offset). prepare-release.yml computed
TODAY via `TZ=UTC date +%Y%m%d`, which is only "correct" if the team
operates on UTC -- but this project's production server logs,
deployment cadence, and team are all Korea Standard Time. Any release
cut between midnight and 9am KST would silently tag itself with
yesterday's date.
Fixed by using `TZ=Asia/Seoul date +%Y%m%d` instead.
Per spec: the sequence number is a per-day counter that resets on
date change and starts at 0, not 1. The first release of a day is
quant_YYYYMMDD.0.hash, the second .1, etc.
Previous commit fixed *counting* today's releases via the Gitea API
(instead of the always-empty local git tags from a shallow checkout),
but still added +1 on top, which would have produced 1, 2, 3... for
the first, second, third releases of a day instead of 0, 1, 2.
DEPLOY_COUNT is now just RELEASES_TODAY directly.
User asked why every release tonight had the same "sequence number"
(quant_20260711.1.*) despite creating three of them. Confirmed via
API: tags b7591fb, 6ab270f, and e49922e all exist for 2026-07-11, all
claiming to be deploy #1.
Root cause: `actions/checkout@v4` (no fetch-depth/fetch-tags options)
does a shallow, tags-less clone by default. Each prepare-release.yml
run happens in a brand-new container, so `git tag -l "quant_${TODAY}.*"`
always sees zero local tags regardless of how many releases actually
exist -- DEPLOY_COUNT was permanently stuck at 0+1=1.
Fix: query GET /repos/{repo}/tags via the Gitea API (same token/curl
pattern already used elsewhere in this workflow) to count today's
actual tags, instead of relying on the job's local, incomplete git
state.
Run #2006 proved the deploy itself is fully working now: checks 1-5
all passed (HTTP 200, login content, CSS, service active, release
verified) -- only check 6 failed, and the log shows exactly why:
❌ [6/6] DB authentication errors found in logs (0
0 occurrences)
`grep -c PATTERN` exits with status 1 whenever the count is 0, even
though it still correctly prints "0" to stdout. The old
`grep -c ... || echo "0"` therefore printed grep's own "0" AND (because
grep's nonzero exit triggered the `||`) a second "0" from the fallback
-- a two-line "0\n0" that can never equal the string "0" in the
subsequent `[ "$DB_ERRORS" = "0" ]` check. So the *healthy* case (zero
DB errors) was the one that always failed this check.
Fixed by using `|| true` instead of `|| echo "0"`: it neutralizes
grep's exit code (needed to avoid an instant abort under `set -e
-o pipefail`, same class of bug as the earlier `git config user.name`
incident) without adding any extra output.
Run #2005's Health Check job hung for 18+ minutes (well past its own
timeout-minutes: 10) instead of failing within seconds. Killed the
zombie container manually via 'docker kill' on the runner host.
Root cause: the pre-fix curl calls to the unreachable
$DEPLOY_HOST:5000 had no --connect-timeout/--max-time, so each of the
20 retry attempts could hang on the OS's default TCP timeout instead
of failing fast; the job-level timeout-minutes didn't reliably cut it
off either (act_runner enforcement gap, not something we control from
the workflow file).
This is now largely moot after the previous commit (health checks run
against 127.0.0.1 on the server itself, where curl returns
near-instantly), but added explicit timeouts everywhere as a second
line of defense against the same failure mode recurring:
- Gitea API curl calls (release fetch, artifact download):
--connect-timeout 10 --max-time 30/120
- Local 127.0.0.1 health-check curls: --connect-timeout 5 --max-time 10
- All ssh/scp invocations: -o ConnectTimeout=10
No single curl or ssh call in this workflow should now be able to
hang indefinitely.
Root cause confirmed by direct test:
curl --connect-timeout 5 http://178.104.200.7:5000/Account/Login -> 000
quantengine.service sets ASPNETCORE_URLS=http://127.0.0.1:5000 (loopback
only, by design -- Nginx is the only public entry point, proxying
quant.taxbaik.com to it). The Gitea Actions runner is not the
production host, so its direct curl to $DEPLOY_HOST:5000 was always
going to hit a closed port. Run #2005 is direct proof: "Deploy to
Production" succeeded, the site was reachable over HTTPS the whole
time, and journalctl was clean -- yet "Health Check & Verification"
burned through all 20 retries (60s) because it was polling the wrong
address entirely. This check has likely never once passed on this
service's actual network layout.
Fix: wrap the HTTP-200 / login-content / CSS retry loop in a single
SSH session that runs curl against 127.0.0.1:5000 on the production
server itself -- consistent with how the service-status and DB-error
checks already correctly run remotely. Removed the redundant
per-attempt SSH round trips for service status (now a plain local
command inside the same remote script) and dropped the separate
"Setup SSH (for service check)" step's curl usage entirely.
Found via SSH log analysis (Run #2004, task 2336): the deploy script's
own echo output revealed the bug directly --
Deploy Dir: /home/kjh2064/deployments/quantengine_$RELEASE_TAG_$COMMIT
tar (child): /tmp/$ARTIFACT: Cannot open: No such file or directory
$ARTIFACT, $RELEASE_TAG, $COMMIT were printed as LITERAL TEXT instead
of their values. Root cause: the heredoc used a quoted delimiter
(<< 'REMOTE'), which correctly prevents the local runner shell from
expanding anything inside it -- but the script still relied on that
expansion happening for these three variables. They were never
actually being passed to the remote bash process at all; this path
had likely never worked.
Fix: pass ARTIFACT/RELEASE_TAG/COMMIT/SERVICE_NAME as env-var
prefixes on the remote `bash -s` invocation (`"VAR='...' bash -s"`),
which the LOCAL shell does expand (since it's a normal double-quoted
string, not part of the quoted heredoc). The heredoc body itself
stays fully remote-evaluated (DEPLOY_HOME=$HOME correctly resolves
to the remote user's home, not the runner's).
Also fixed: COMMIT was being read from the release's
`target_commitish` field, which is the branch name the tag points to
("main"), not a commit SHA -- confirmed by the same log ("Commit:
$COMMIT" would have printed "main" once the heredoc bug was fixed).
Since our tags are always "quant_YYYYMMDD.count.hash"
(prepare-release.yml), the hash is now parsed directly out of the
tag name instead.
Found via SSH log analysis (Run #2003, task 2334): the "Verify SSH
Key and Secrets" step failed immediately with
"DEPLOY_SSH_KEY_B64 or DEPLOY_SSH_KEY not configured" -- both were
empty. Queried GET /repos/{repo}/actions/secrets directly and found
the actually-registered secrets are named SSH_PRIVATE_KEY and
QUANTENGINE_DB_PASSWORD; DEPLOY_SSH_KEY_B64/DEPLOY_SSH_KEY were never
created, despite CLAUDE.md claiming "SSH credentials: SSH_KEY
registered in Gitea Secrets".
Every past deploy-prod.yml run that reached the SSH step (e.g. Run
#1991's Pre-Deployment Verification) failed here for the same reason
-- this was never a working path, just never diagnosed down to the
secret name before now.
Fix: check secrets.SSH_PRIVATE_KEY first (with the same PEM-vs-base64
auto-detection used for the legacy names), falling back to
DEPLOY_SSH_KEY_B64 / DEPLOY_SSH_KEY in case those get added later.
Applied to all three places that build ~/.ssh/deploy_key (deploy job
verify + setup, and post-deploy-check's setup).
kjh2064