# Monitoring — Open Items

Only keep unfinished monitoring/observability work here.

## Recently resolved / verified

- `/etc/zlh-monitor` is now the runtime source of truth.
  - Prometheus runs from `/etc/zlh-monitor/prometheus/prometheus.yml` via `/etc/default/prometheus`.
  - Grafana provisions from `/etc/zlh-monitor/grafana/provisioning` via `/etc/default/grafana-server`.
  - `/etc/zlh-monitor/README.md` describes the new single-source layout.

- Game/dev dynamic discovery empty-state now works.
  - `/sd/exporters -> []`
  - `/monitoring/game-dev -> containers: []`
  - monitor-side `ALLOW_EMPTY="true"` added in `/etc/zlh-monitor/game-dev-discovery.env`
  - one-shot sync succeeded and wrote `0 targets` to `/etc/zlh-monitor/prometheus/file_sd/game-dev-alloy.json`
  - Prometheus refreshed and stale `game-dev-alloy` VMIDs `5205` and `6084` disappeared

- Prometheus runtime config is clean.
  - validates with `promtool`
  - active jobs are only:
    - `prometheus`
    - `alloy -> 127.0.0.1:12345`
    - `game-dev-alloy -> /etc/zlh-monitor/prometheus/file_sd/game-dev-alloy.json`
  - no active `node_exporter` job
  - no active cAdvisor job
  - no active `:9100` jobs

- Remaining stale/static Prometheus targets are removed from active runtime config.
  - `alloy-dev-6076` no longer appears as an active runtime target
  - `cadvisor 10.60.0.11:8081` no longer appears as an active runtime target
  - `zerolaghub-nodes 10.60.0.19:9100` no longer appears as an active runtime target

- Grafana provisioning is live.
  - source-managed datasource:
    - `uid=prometheus`, `name=Prometheus`, `url=http://localhost:9090`, `default=1`
  - provisioned dashboards:
    - `Core Service - Detail`
    - `Core Services - Overview`
    - `Dev Containers`
    - `Game Containers`
  - legacy node_exporter router dashboard archived to `/etc/zlh-monitor/grafana/dashboards-archive/core-routers.json`

- Important game/dev metrics path works through remote-write.
  - API discovery emitted both test servers:
    - `6085 / dev-6085 / 10.100.0.23 / dev`
    - `5206 / mc-vanilla-5206 / 10.200.0.79 / game`
  - file_sd sync wrote both targets and Prometheus discovered both
  - container Alloy remote-write works after Prometheus was rebound to `10.60.0.25:9090`
  - dashboard-style metrics exist for both:
    - `node_uname_info`
    - CPU
    - memory
  - observed values:
    - `6085 dev-6085 dev-container CPU ~1.04% Mem ~6.58%`
    - `5206 mc-vanilla-5206 game-container CPU ~0.68% Mem ~61.69%`

- Current monitoring model clarified.
  - Not API-only: API owns discovery/metadata.
  - Canonical metrics path is container Alloy remote-write to Prometheus.
  - `game-dev-alloy` scrape-up panels should not be relied on while container Alloy listens on `127.0.0.1:12345`.
  - Direct scrape health would require intentionally exposing container Alloy HTTP on a reachable address such as `0.0.0.0:12345` plus security review.

- Services active after cleanup:
  - `prometheus`
  - `grafana-server`
  - `alloy`
  - discovery timer

## Launch blockers

1. Finish network exposure verification/lockdown.
   - Prometheus now listens on `10.60.0.25:9090` for container remote-write and self-scrape.
   - Grafana remains loopback-only on `127.0.0.1:3000`.
   - zlh-monitor local Alloy remains loopback-only on `127.0.0.1:12345`.
   - local node_exporter is disabled.
   - Required validation: Prometheus is reachable only from intended container remote-write and trusted admin/monitoring paths; Grafana is not externally exposed; node_exporter is not listening or is firewalled; host firewall/nftables posture is acceptable.

2. Add API health scrape or equivalent application-level health signal.
   - Observed `/health` previously returned `404 Cannot GET /health`.
   - Acceptance: Prometheus has an API app health target or equivalent application-level health signal.

3. Add launch-debug lifecycle telemetry.
   - Missing/incomplete: `ready`, `connectable`, `operationInProgress`, `operationType`, backup/restore state, code-server state.
   - Acceptance: dashboards or Prometheus queries can answer what a game/dev container is doing and why it is not ready.

4. Add centralized logs.
   - No Loki, Promtail, or log-ingesting Alloy config found in prior audit.
   - Acceptance: API, Agent, provisioning, backup/restore, delete/teardown, Velocity/DNS/edge, and discovery-sync logs are queryable from the monitoring surface.

5. Tighten bearer token storage.
   - Token previously appeared in Prometheus YAML, backup configs, and `/etc/zlh-monitor/game-dev-discovery.env`.
   - Acceptance: token stored in tight-permission env file or systemd credentials; not world-readable; not leaked in labels/URLs/logs.

6. Finish lifecycle smoke test delete step.
   - Delete test dev/game servers.
   - Confirm API discovery drops them.
   - Confirm file_sd drops them.
   - Confirm dashboard metrics age out as expected.

## Standardization work

- Standardize labels: prefer canonical `container_type` over mixed `ctype` / `container_type`.
- Decide canonical label spelling for hostname/name/server identity.
- Keep required game/dev labels minimal: `vmid`, `instance`, `container_type`.
- Keep optional metadata useful but not noisy.
- Review any user/customer-identifying labels before dashboard or alert exposure.

## Alerts to add

- discovery sync failed
- file_sd file stale / not updated recently
- missing/stale remote-write metrics by VMID/type
- API `/sd/exporters` failure
- API app health failure
- Prometheus scrape failures by important job
- Grafana dashboard provisioning missing/failure
- log pipeline down once Loki/Alloy logs exist

## Follow-up improvements

- Add Loki or Alloy log collection for API, Agent, provisioning, backup/restore, Velocity/DNS/edge actions.
- Add panels for lifecycle state: `ready`, `connectable`, `operationInProgress`, `operationType`, backup/restore status, code-server status.
- Add dashboard row/panels for stale/deleted targets.
- Add a runbook for discovery sync failure.
- Add a runbook for failed game/dev provisioning debugging from monitoring only.