fix(grafana): no false DATA SOURCE ISSUE page when influxdb restarts (#1682)

[groupsky]

Fixes #1682.

Why a page was sent although execErrState: OK

Grafana 9.5 logs from the 2026-10-11 deploy:

1. 11:00:00 tick: InfluxDB was still opening its store, so the queries failed. execErrState: OK turned this into Normal. No page came from this step.
2. InfluxDB began to answer at 11:00:04. The heat pump query (current_power, phase B, last 2 min) then ran without error but got no rows. The writers had not yet back-filled the minutes when InfluxDB was down. The data is there now, but it was not there at that moment.
3. noDataState: NoData put the rule into the NoData state. In Grafana 9.5, NoData does not wait for for:. It is sent at once as an alert named DatasourceNoData, and telegram-bridge prints that as "📵 DATA SOURCE ISSUE" (docker/telegram-bridge/src/message-utils.js).
4. Both "Термопомпа не работи" (old UI rule 3A5nIop4k) and "Термопомпа не работи (Webhook)" did this, so 2 pages came.

Changes

1. grafana waits for influxdb: condition: service_healthy. InfluxDB already has a real healthcheck: /health on port 8086. That port opens only after the store is open, and the 300 s start period covers slow opens. This is the smallest way to do it; a wait loop in Grafana's entrypoint would do the same job with more code. Limits:
   • The wait works only when compose starts both in one up, as a deploy that recreates both does. After a reboot, Docker's restart policy does not wait. Change 2 covers that case for the heat pump rule.
   • deploy.sh uses up --no-deps. If only grafana changes, compose v2.18 does not wait. If both change, it waits up to ~300 s, which is inside COMPOSE_TIMEOUT=600.
   • The infrastructure CI test now starts grafana ha with --no-deps. That test needs only the ingress network, and InfluxDB cannot start in that job (its bind-mounted data dir is owned by root). It could not start before this PR either; nobody saw it because nothing waited for it.
2. The heat pump "not working" rule gets noDataState: Alerting. In 9.5 the Alerting state does wait for for: 15m. A short no-data gap after a restart stays silent. If no rows come back for more than 15 min, the rule still pages under its own title. A full InfluxDB outage, where queries fail, is still seen through the rules that have execErrState: Alerting (boiler controller, sunseeker). Change 1 alone is not enough: the empty window comes from the writers back-filling late, not from Grafana starting early.
3. 3 old UI-made rules are deleted with a deleteRules file (same way as delete-boiler-manual-mode-alert.yaml). They copy provisioned rules, so each alert was sent twice. Before the delete, all 9 rules in that folder were exported on the host (full rows from grafana.db, plus a copy of the DB file). There is no Grafana admin login on the host, so the Grafana API could not be used.

Old "General Alerting" rules (UI-made, not in the repo)

Contact point: every rule below, and every twin, has no labels that match a route, so all of them use the default telegram-webhook.

Legacy uid	Title	Provisioned twin	Query	Threshold	for	Contact	Action
3A5nIop4k	Термопомпа не работи	webhook-3A5nIop4k	same (*)	same (<30 W)	same 15m	same	deleted
lA5nSTtVk	Main Power alert	webhook-lA5nSTtVk	same	same (>10 kW)	same 5m	same	deleted
z157Sot4z	Water pump power alert	webhook-z157Sot4z	same	same (>1 kW)	same 15m	same	deleted
xjXziW_4k	Термопомпа претоварена	(webhook-lA5nSTt4k is close)	15 min window, not 10 min	>1000 W, not >3000 W	15m, not 10m	same	kept
3057STtVkz	AC alert	webhook-3057STtVkz	other (min/max, not median)	210/248 V, not 207/253 V	0, not 4m	same	kept
RJcnSop4z	External alert	none	–	–	5m	default	kept
rA5nSotVkz	Temperature alert	none	–	–	5m	default	kept
j05nSotVz	Humidity alert	none	–	–	5m	default	kept
eAc7ITtVk	Output alert	none	–	–	5m	default	kept

(*) The only differences: the expression step uses the old datasource id -100 instead of __expr__ (both mean "expression"), and 3A5nIop4k keeps an unused raw-query text from an old editor mode (rawQuery: false).

For the owner: the 6 kept rules are still only in the Grafana DB. xjXziW_4k alerts at >1000 W but its message says 3000 W, and it keeps noDataState: NoData.

Not changed (said here instead of built)

Other provisioned rules still use noDataState: NoData. The only one with a short window is Main Power (Webhook), with a 5 min window. If InfluxDB is killed and needs ~200 s to open, that window could also be empty for a moment. Its Alerting state means ">10 kW", so noDataState: Alerting would be a wrong page there. If it ever fires on a deploy, it needs its own fix.

Done-when proof

The deploy check (both containers recreated, then the Telegram send log) will be added as a comment on #1682 after merge.