The estate had zero alert rules. The only contact point was Grafana's factory
default pointing at the literal string <example@email.com>, and beta's
untracked compose override routed Grafana's SMTP at prod's Postfix — so
alerts, had any existed, would have been delivered by the box most likely to
be on fire. Prod served 86 5xx in 24h including five /healthz failures and
nobody was told.
Alerting (deploys to beta, which is where Grafana runs):
- 12 Loki-based rules. There is no Prometheus in this estate and Loki is the
only datasource, so every rule is log-derived.
- Thresholds come from a 24h backtest that happens to contain a real ~20min
prod outage at 04:20Z. Absolute counts, not ratios: prod runs ~7 req/min, so
one bad request is 1.4% and a ratio alert would scream all night. The 5xx
burst rule fires on the outage's 45 and 22 buckets and on nothing else in
the day; the largest benign bucket all day was 5.
- Routed to a new private #ops-alerts channel, not to any existing channel —
#weather-events, #announcements and #prod are product surfaces that notify
real subscribers.
- AlertingWatchdog is a dead-man's switch; its value is its absence.
- Delivery is proven, not assumed: the rules were provisioned into a throwaway
Grafana against beta's live Loki and a real alert arrived in Discord. This
matters because GET /api/v1/provisioning/contact-points returns [REDACTED]
for the URL — a contact point holding an uninterpolated env var looks
perfectly healthy and pages nobody. The only proof is a message arriving.
CI gains a structural check, because the existing one only proves YAML parses:
an alert rule whose condition names a missing refId is valid YAML, provisions
cleanly, and never fires. It also hard-fails on a literal Discord webhook in
the repo. Verified against all three breakages deliberately introduced.
Postfix supervision:
- The 13h outage was a boot-ordering race, not a Docker renumbering: postfix
started at 08:09:49, wg0 came up at :51, postfix fataled at :52 on a missing
docker_gwbridge address, and dockerd did not finish starting until 08:10:16.
Stock postfix@.service is ordered only After=network-online.target and ships
no Restart=, so one lost race became a permanent outage.
- An ExecStartPre gate now blocks up to 60s until every inet_interfaces
address actually exists, which absorbs the transient case inside a single
start attempt. That makes bounded retry correct: 5 attempts in 600s, then
failed — a genuinely broken config reaches a visible failed state in ~100s
instead of re-fataling every 15s forever.
- A 5-minute watchdog timer retries indefinitely and runs reset-failed, so
"failed" still self-heals. Worst case is ~5 minutes, not 13 hours.
- Wants=, not Requires=: a dockerd failure must not take down the loopback and
mesh listeners that do not depend on Docker at all.
Health checks must read config with `postconf -c`, never postmulti/postqueue/
postfix — those three RESOLVE inet_interfaces and so fatal precisely when an
address is missing, which made the first version of this check report
status=ok bound=0/0. A health check that fails open is worse than none.
DEPLOY.md carries the monitoring contract, including that systemctl is-active
postfix is a known-false signal: postfix.service is a wrapper whose
ExecStart=/bin/true, so it reports active forever while the real postfix@-
instance is failed with zero listeners. Reproduced live.
Known gap, documented not fixed: Alloy ships Docker stdout, Caddy files and
app JSONL, not journald — so no Postfix line reaches Loki and the mail rules
cannot fire until loki.source.journal is added.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
main took two direct PRs (#19 shell-lint, #21 the thermograph-daemon Go
service) while dev accumulated the ERA5 lake stack; both sides added compose/
stack services and touched the same seams. Resolutions keep both worlds:
- compose + Swarm stack: lake AND daemon are sibling services; backend env
carries THERMOGRAPH_LAKE_URL and THERMOGRAPH_INTERNAL_TOKEN.
- deploy.sh: backend deploys roll (backend lake daemon); all adds frontend.
- requirements.txt: websockets/apscheduler stay removed (moved to the Go
daemon), duckdb stays (the lake role's engine).
Merged tree: backend suite green, both compose configs validate.
The gateway bot and APScheduler were long-lived stateful I/O loops running
inside the async web app under a leader election. They move into a single Go
binary that owns ONLY that I/O -- websocket, RESUME, heartbeat, backoff, timers.
It owns no grading logic. Anything needing data calls back over a new
internal-only surface (/internal/discord/grade, /internal/jobs/*). Grading
depends on polars and the parquet cache; reimplementing it in Go would let the
bot's grades drift from the API's. The grade route returns gateway-ready JSON
and Go relays the bytes verbatim.
The binary ships in the backend image and runs as a second compose service off
the same tag, so the two ends of the /internal/* contract can never skew.
deploy.sh rolls daemon alongside backend -- without that the service would never
be created, since a single-service deploy uses --no-deps. It also probes the
image first and skips the daemon when rolling a tag that predates the binary:
infra tracks main while image tags are env-staged, so a host can legitimately be
asked to roll an older backend image, and creating the service anyway would
leave a container crash-looping on a missing binary.
replicas: 1 with order: stop-first replaces the leader election -- Discord
permits one gateway connection per bot token.
THERMOGRAPH_INTERNAL_TOKEN is optional: both ends derive it from
THERMOGRAPH_AUTH_SECRET via HMAC under a domain-separation label, so this needs
no new vault entry. The derivation is pinned to a shared cross-language test
vector asserted on both sides, so drift fails CI instead of 401ing every call.
Fail closed when neither secret is set.
Improvements over the Python: a close intended for RESUME uses 4000 rather than
1000 (Discord invalidates a session closed 1000, so the old default defeated its
own resume); MESSAGE_CREATE runs on a bounded worker pool; and a malformed HELLO
returns an error rather than a clean reconnect, which would otherwise reset
backoff and hot-loop against the gateway.
365 Python tests pass; Go build/vet/test -race clean; shellcheck 0 findings.
Adds .forgejo/workflows/shell-lint.yml (pinned shellcheck v0.11.0 + sha256, -x,
default severity, fail on any finding, not path-filtered) and drives all 26
scripts to zero findings.
Two defects shellcheck cannot see:
render-secrets.sh left DECRYPTED vault contents in /tmp whenever a sops decrypt
failed -- the caller's set -e aborted the function before either cleanup ran.
Now removed on every exit path, with `|| return 1` on both sops calls so a
decrypt failure can never write a partial /etc/thermograph.env regardless of the
caller's shell options. Explicitly not a `trap ... RETURN`: such a trap set in a
sourced function persists into the caller's shell and re-fires when the caller's
next `.`/source completes, where the function-local tmp is unset -- fatal and
silent under deploy.sh's set -u. The file now records that reasoning.
autoscale.sh ran `set -eu` without pipefail while piping docker stats into awk,
so a failed left side was swallowed and the loop autoscaled on empty input.
Promoted to pipefail with a missed sample treated as a skip; verified busybox ash
in docker:27-cli supports it.
Also: capture-fixtures.sh's `jq . || cat` ran cat after jq had consumed stdin,
silently writing truncated fixtures; deploy.sh/deploy-stack.sh `# shellcheck
source=` paths corrected for the monorepo layout.
Prod DB dumps and Forgejo (db + data volume) are now copied off-box to the
era5-thermograph S3 bucket, age-encrypted to the vault recipient so the host
age key at /etc/thermograph/age.key decrypts them for restore. Nothing plaintext
leaves the box; no local intermediate (streamed via age | rclone rcat).
ops-cron: the prod pg_dump job gains an off-box push; a new forgejo-backup job
dumps Forgejo's Postgres + tars its data volume on beta (the git host had NO
backup at all). 30-day off-box retention on both. S3 creds are Forgejo Actions
secrets (S3_ENDPOINT/BUCKET/ACCESS_KEY/SECRET_KEY), mirrored into the SOPS vault
(prod.yaml/beta.yaml) for host-side use. Contabo needs path-style addressing.
Adds deploy/backup/README.md with the DR restore runbook.
The cutover moved the host checkout's deploy tree to /opt/thermograph/infra/deploy/,
but the stack file's absolute bind sources (db init, env-entrypoint.sh, autoscale.sh)
still pointed at the old /opt/thermograph/deploy/ -- a service update kept the stale
mounts and new tasks failed with 'bind source path does not exist'.
/opt/thermograph becomes a monorepo checkout: deploy.sh gains INFRA_DIR (git
ops at the root, compose work cd'd into infra/), lock/tags/render paths move
under infra/deploy/, image-path defaults become emi/thermograph/backend|
frontend across compose, stack, and the tag-prune. docker-compose.yml pins
name: thermograph -- without it compose run from .../infra derives project
"infra" and recreates the whole stack beside the running one;
deploy-dev.sh pins COMPOSE_PROJECT_NAME=thermograph-dev (env wins over the
file key) to keep LAN dev's separate project.
One root workflow set replaces the four repos' copies (deleted -- root-only
is where Forgejo reads them, and dead copies are a trap): per-domain
build-push with explicit image paths (emi/thermograph/backend|frontend; the
old github.repository-derived path collides in a monorepo), path-filtered
per-domain beta/prod/dev deploys, a domain-input reusable build check, a
single always-reporting PR gate (path-filtered required checks deadlock
auto-merge), a new infra-sync pipeline (host checkout + secrets render on
infra/** pushes), and ports of secrets-guard / ops-cron /
observability-validate to monorepo paths.
