# 15 — Observability

## Summary

Production has live metrics and tracing infrastructure as of 2026-04-25.
A self-hosted **VictoriaMetrics + Jaeger + Grafana** stack runs on
`88oakappsUpdate` (Linode VPS, also home to the self-hosted PostHog
deployment). A `vmagent` sidecar in the honeyDue k3s namespace scrapes
the api Pods' `/metrics` endpoint every 15 seconds and remote-writes to
`https://obs.88oakapps.com/api/v1/write`. Grafana is at
`https://grafana.88oakapps.com` with a pre-provisioned RED dashboard.

What we still don't have: log aggregation (Dozzle and `kubectl logs`
fill the niche for now), alerting (no PagerDuty/Slack on errors), and
full distributed tracing (OTel SDK is wired in app code but app-side
instrumentation beyond HTTP routes hasn't shipped yet).

The whole observability stack costs **$0** incremental and uses ~700 MB
RAM on `88oakappsUpdate` (5% of its free RAM). It runs as a separate
docker-compose project from PostHog so neither product's lifecycle
touches the other.

## What we have

### 1. Metrics — VictoriaMetrics + vmagent

```
honeyDue k3s (Hetzner)                   88oakappsUpdate (Linode)
┌───────────────────────────┐            ┌──────────────────────────┐
│ api Pods (3) :8000/metrics│            │ /opt/honeydue-obs/       │
│   prometheus/client_golang│            │ ┌──────────────────┐     │
│                           │            │ │ VictoriaMetrics  │     │
│ vmagent ──── scrape 15s   │            │ │  30d retention   │     │
│         remote_write ─────┼────────────┼─→ /api/v1/write   │     │
│         (HTTPS, bearer)   │            │ │  (mem 256 MB)    │     │
└───────────────────────────┘            │ └──────────────────┘     │
                                          └──────────────────────────┘
```

The Go API exposes `/metrics` in Prometheus exposition format. Histograms
are defined in `internal/prom/metrics.go` and registered globally:

| Metric | Labels | Source |
|---|---|---|
| `http_request_duration_seconds` | `route, method, status` | Echo middleware around every handler |
| `gorm_query_duration_seconds` | `table, operation` | GORM before/after callbacks (no ctx threading needed) |
| `b2_upload_duration_seconds` | `bucket, result` | Wrapped `s.backend.Write` in `internal/services/storage_service.go` |
| `b2_upload_bytes_total` | `bucket, result` | Counter alongside the duration histogram |
| `apns_send_duration_seconds` | `result` (`ok`/`bad_token`/`error`) | Wrapped APNs `PushWithContext` in `internal/push/apns.go` |
| `fcm_send_duration_seconds` | `result` | Wrapped FCM HTTP v1 send in `internal/push/fcm.go` |
| `asynq_job_duration_seconds` | `task_type, result` | Histograms registered; middleware not yet attached (Step 3) |
| `go_*`, `process_*` | (standard) | `prometheus/client_golang/prometheus/collectors` defaults |

The previous custom monitoring at `/metrics` was renamed to
`/metrics/legacy` so the canonical `/metrics` emits proper histograms
suitable for `histogram_quantile()` rollups. The legacy endpoint stays
because the GoAdmin dashboard reads it.

#### vmagent in k3s

Lives at `deploy-k3s/manifests/observability/vmagent.yaml`. One replica,
`mem_limit: 256Mi`, scrapes by Kubernetes pod-discovery filtered to
`app.kubernetes.io/name=api` and remote-writes to
`https://obs.88oakapps.com/api/v1/write` with a bearer token from
`OBS_INGEST_TOKEN` in `deploy/prod.env` (substituted into a Secret at
deploy time).

The agent buffers locally to `/tmp/vmagent` (emptyDir, 512 MB cap), so
brief obs outages don't drop samples. Persistent queue replays on
reconnect.

NetworkPolicies in the honeydue namespace allow egress from vmagent to:
- DNS (kube-dns / coredns)
- Kubernetes API (`10.43.0.0/16:443`) for pod discovery
- api Pods on `10.42.0.0/16:8000`
- The public obs endpoint over `0.0.0.0/0:443`

These are scoped tight — vmagent can't reach Postgres, Redis, B2, or
any other external service.

### 2. Tracing — Jaeger all-in-one

Jaeger 1.62 with badger storage runs alongside VictoriaMetrics. The
collector accepts:
- OTLP/HTTP at `https://obs.88oakapps.com/v1/traces` (bearer-token gated)
- OTLP/gRPC at `:4317` (localhost-only)
- Native Jaeger protocols at `:14268` etc. (localhost-only)

Retention: ~7 days at current scale before badger rotates. UI at
`https://grafana.88oakapps.com` via the Jaeger datasource.

**Status of app-side instrumentation**: the histograms are populating
metrics. The OTel exporter wiring in `cmd/api/main.go` is **not yet
shipped**. When it does ship, every `POST /api/auth/login/` will produce
a flame-graph trace with HTTP → handler → SQL → B2 → APNs spans.
Tracking issue: gitea#3.

### 3. Dashboards — Grafana

`https://grafana.88oakapps.com` (Cloudflare-fronted, basic auth via
Grafana itself, admin credentials in `deploy/prod.env`).

Datasources auto-provisioned at container startup from
`/opt/honeydue-obs/data/grafana-provisioning/datasources/datasources.yaml`:
- VictoriaMetrics (Prometheus type, `http://victoriametrics:8428` in-network)
- Jaeger (`http://jaeger:16686` in-network)

Pre-provisioned dashboard: `honeyDue API — RED` at
`/d/honeydue-red`. Top row uses the legacy custom metrics
(`http_endpoint_requests_total`, `http_requests_total`) which started
flowing the moment vmagent attached. Lower rows use the new histograms
(`http_request_duration_seconds_bucket` p50/p95/p99 by route, GORM p95
by table, B2 upload p95, APNs/FCM send p95, Go memory + goroutines).
Lower rows populated immediately after the api rebuild that shipped
`internal/prom`.

### 4. `kubectl logs`

Every container's stdout/stderr is captured by containerd and readable
via kubectl:

```bash
# Live tail from all api pods
kubectl logs -n honeydue -l app.kubernetes.io/name=api -f --prefix

# Last 100 lines
kubectl logs -n honeydue -l app.kubernetes.io/name=api --tail=100

# Previous pod's logs (before the most recent restart)
kubectl logs -n honeydue <pod-name> --previous

# Events (not logs — k8s-level state changes)
kubectl get events -n honeydue --sort-by=.lastTimestamp
```

**Retention**: containerd rotates logs when they exceed 10 MB (default).
Only the last ~20 MB of logs is retained per container, on-disk on the
node. Once a pod is deleted, its logs are gone.

For persistent log access we'd need aggregation (see §What we still
don't have).

### 5. `kubectl top`

Pod and node resource usage via metrics-server:

```bash
kubectl top nodes
# NAME                CPU(cores)   CPU(%)   MEMORY(bytes)   MEMORY(%)
# ubuntu-8gb-nbg1-1   169m         4%       748Mi           9%

kubectl top pods -n honeydue
```

In-memory only; last few minutes of data. For historical trends use
the Grafana dashboard, which exposes the same data via the `go_*` and
`container_*` (kubelet cAdvisor) metrics.

### 6. Cloudflare Analytics

CF Dashboard → Analytics & Logs. Per-zone aggregate stats:
requests/sec, bandwidth, cache hit ratio, top status codes, top paths,
bot traffic score. Good for spotting macro trends ("suddenly 10× more
502s today") that wouldn't show up in a single-pod sample.

Free tier retention: 7 days of aggregate stats.

### 7. Neon dashboard

Neon console → project → Monitoring: compute utilization (CU-hours),
slow queries, active connections, storage usage. Useful for "is the
DB busy?" and free-tier limit watching. The new
`gorm_query_duration_seconds` histogram covers the application side
of the same question with much better latency tail visibility.

### 8. Kubernetes events

`kubectl get events` shows cluster-level state changes: pod scheduling,
failures, image pulls, probe failures. Useful for post-mortem on
deploys.

Retention: events are stored in etcd but default to 1 hour.

## What we still don't have

### No log aggregation

Individual pod logs are on the node. For multi-pod debugging ("show me
all api pod logs for user X") we have to:

```bash
# Query all at once with stern (if installed)
stern -n honeydue api

# Or per-pod
kubectl logs -n honeydue <pod> | grep user_id=12345
```

This works but doesn't scale across many pods.

**What we'd add**: [Loki](https://grafana.com/oss/loki/) on
`88oakappsUpdate` next to the existing obs stack. Adds ~512 MB RAM
plus a Promtail (or Vector/Alloy) DaemonSet in k3s. Defer until log
search becomes a recurring pain point — `stern` + `grep` is fine at
current pod count.

### No alerting

No PagerDuty, no Slack webhooks, no email on "api is returning 500s."
The operator finds out when users complain.

Cheapest fix path:
1. Grafana alerting (built into Grafana 11) — alert rules over the
   existing histograms (e.g., `histogram_quantile(0.95, ...) > 1s`).
   Routes to Slack via webhook. **Zero infra cost.**
2. [Uptime Kuma](https://github.com/louislam/uptime-kuma) on
   `88oakappsUpdate` — pings `/api/health/` from outside the cluster
   every minute; complements the in-cluster view.

We'd want both eventually. Grafana alerting first because the data is
already there.

### Distributed tracing — fully integrated

The OTel SDK is wired in `cmd/api/main.go` and `cmd/worker/main.go` and
ships traces to Jaeger via `obs.88oakapps.com/v1/traces`. Every public
service method now takes `ctx context.Context` and routes its SQL through
`repo.WithContext(ctx)`, which means **every authenticated API endpoint
produces a fully-nested flame graph** in Jaeger.

| Span source | Status |
|---|---|
| `otelecho.Middleware` — span per HTTP request | ✅ live |
| Auth middleware DB lookups (`m.db.WithContext(ctx)`) | ✅ live |
| All repos via `repo.WithContext(ctx)` (`otelgorm` plugin) | ✅ live |
| Manual span around `storage_service.Upload` (B2 PutObject) | ✅ live |
| Manual span around APNs `Send` / `SendWithCategory` | ✅ live |
| Manual span around FCM `sendOne` | ✅ live |
| Asynq middleware — span per task type with retry/payload attrs | ✅ live |

Migrated services (every public method takes ctx):
- `AuthService` — login, register, refresh, logout, me, verify-email,
  forgot/reset-password, update-profile
- `TaskService` — all 25+ task and completion methods
- `ResidenceService` — all 15 methods including share-codes
- `ContractorService` — all 9 methods
- `DocumentService` — all 10 methods
- `NotificationService` — all 12 methods
- `SubscriptionService` — all 12 methods including Apple/Google IAP

Sample trace for `GET /api/tasks/` (warm cache, post-optimization):

```
GET /api/tasks/                                              (229ms)
└── service: SELECT * FROM task_task WHERE residence_id IN
              (SELECT id FROM residence_residence WHERE...)   (227ms)
```

Two spans total. The auth path runs entirely from Redis + in-memory
cache (zero SQL queries) thanks to the 1-hour token TTL and 5-min user
TTL. The residence-ID lookup is folded into the tasks query as a
Postgres subquery, so a single network round-trip to Neon services the
whole request. See Chapter 8 §"Optimizations layered on top" for the
optimization stack.

Earlier trace, before the optimization stack landed (commit 88fb175):

```
GET /api/tasks/                                              (2473ms)
├── auth: SELECT * FROM user_authtoken WHERE key=...           (1506ms)
├── auth: SELECT * FROM auth_user WHERE id=7                    (333ms)
├── service: SELECT id FROM residence_residence WHERE...        (736ms)
└── service: SELECT * FROM task_task WHERE residence_id IN(...) (226ms)
```

10× improvement from 2,473ms to 229ms by cutting query count
(5 SQL → 1 SQL on warm cache). The 227ms in the surviving query is
**1 transatlantic round-trip** to Neon us-east-1 from Hetzner
Nuremberg — the physical floor on the current setup. Eliminated by
migrating Neon to a EU region; tracked in [Chapter 18 §migration
triggers](./18-cost.md) and `docs/observability-plan.md`.

**Migration pattern (for any future services or middleware):** add
`ctx context.Context` as the first arg, change the handler call site
to pass `c.Request().Context()`, and replace `s.repo.X(...)` with
`s.repo.WithContext(ctx).X(...)`. Tests pass `context.Background()`.

### No APM (Application Performance Monitoring)

No continuous profiling. We can answer "which endpoint has the highest
p99 latency?" from the histograms, but not "where in the call stack is
the time going?" without ad-hoc `pprof` runs.

If/when needed: Grafana Pyroscope is the OSS continuous profiler that
fits our stack. Adds ~512 MB RAM. Defer until a CPU performance
incident shows up.

## The app's logging conventions

The Go app uses zerolog and emits structured JSON:

```json
{
  "level": "info",
  "time": "2026-04-24T05:29:40Z",
  "caller": "/app/cmd/api/main.go:189",
  "addr": ":8000",
  "message": "HTTP server listening"
}
```

Log levels: `debug`, `info`, `warn`, `error`, `fatal`. Controlled by
`DEBUG=true|false` in the ConfigMap (true sets level to debug, false
sets level to info).

Every request is logged with method, path, status, request_id, user_id
(if authenticated), latency. Queryable by grep today; ready to ingest
into Loki when we add it.

## Health endpoints

Each service exposes a health endpoint:

| Service | Endpoint | What it checks |
|---|---|---|
| api | `/api/health/` | Process alive (doesn't verify DB) |
| api | `/api/health/live` | Process alive |
| admin | `/` | Next.js is up |
| worker | (none public) | Internal Asynq status |
| api | `/metrics` | Prometheus exposition (vmagent scrapes here) |
| api | `/metrics/legacy` | Custom monitoring metrics for GoAdmin |

Health endpoints are **shallow** — they return 200 if the process is
running and listening. They don't try to reach Postgres/Redis/etc.
Rationale: if Postgres is briefly down, we don't want all api pods to
start failing liveness and cascade-restart.

## obs.88oakapps.com — the ingest endpoint

Public hostname for cross-cluster metric and trace ingest. Cloudflare
in front, nginx on `88oakappsUpdate` enforces a bearer-token check
before forwarding to the local VM/Jaeger containers.

| Path | Forwards to | Purpose |
|---|---|---|
| `/api/v1/write` | `http://127.0.0.1:8428` | Prometheus remote-write (vmagent → VM) |
| `/v1/traces` | `http://127.0.0.1:4318/v1/traces` | OTLP/HTTP traces (app → Jaeger) |
| `/health` | (returns 200) | Reachability probe — also requires auth |
| anything else | 404 | |

Token lives at `/etc/honeydue-obs/secrets.env` (mode 0600 on the box)
and at `OBS_INGEST_TOKEN=` in `deploy/prod.env` (gitignored). To rotate:
generate a new value, update both ends, restart vmagent.

```bash
# Operator: rotate the bearer token
NEW=$(openssl rand -hex 32)
ssh 88oakappsUpdate "sudo sed -i 's|OBS_INGEST_TOKEN=.*|OBS_INGEST_TOKEN=$NEW|' /etc/honeydue-obs/secrets.env"
ssh 88oakappsUpdate "sudo sed -i 's|Bearer [a-f0-9]\{64\}|Bearer $NEW|' /etc/nginx/sites-available/obs.88oakapps.com && sudo nginx -s reload"
sed -i.bak "s|^OBS_INGEST_TOKEN=.*|OBS_INGEST_TOKEN=$NEW|" deploy/prod.env
KUBECONFIG=~/.kube/honeydue.yaml kubectl -n honeydue create secret generic vmagent-remote-write \
  --from-literal=bearer_token=$NEW --dry-run=client -o yaml | kubectl apply -f -
KUBECONFIG=~/.kube/honeydue.yaml kubectl -n honeydue rollout restart deploy/vmagent
```

## Resource budget

| Service | mem_limit | Disk | Retention |
|---|---|---|---|
| VictoriaMetrics | 256 MB | 10 GB | 30 days |
| Jaeger all-in-one (badger) | 256 MB | 10 GB | ~7 days |
| Grafana OSS | 256 MB | 1 GB | — |
| vmagent (in k3s) | 256 MB | 512 MB emptyDir | — |
| **Total** | **~1 GB hard cap** | **~21 GB** | |

Resident usage at idle is much lower (~90 MB on the obs side, ~30 MB
for vmagent). Hard limits exist so a memory leak in any one component
can't squeeze the cohabiting PostHog stack on `88oakappsUpdate`.

## Operator cheat sheet

```bash
# Tail all logs in the namespace
kubectl logs -n honeydue --all-containers=true --tail=50 -l app.kubernetes.io/part-of=honeydue

# Scrape state from vmagent self-metrics
kubectl -n honeydue exec deploy/vmagent -- wget -qO- http://127.0.0.1:8429/metrics \
  | grep -E "scrapes_total|targets|remotewrite"

# Force vmagent to reload scrape config
kubectl -n honeydue rollout restart deploy/vmagent

# Query VictoriaMetrics directly (PromQL)
ssh 88oakappsUpdate 'curl -s "http://127.0.0.1:8428/api/v1/query?query=histogram_quantile(0.95,sum%20by%20(route,le)(rate(http_request_duration_seconds_bucket%5B5m%5D)))" | python3 -m json.tool'

# Restart the obs stack on 88oakappsUpdate
ssh 88oakappsUpdate 'cd /opt/honeydue-obs && sudo docker compose restart'

# Live obs container memory
ssh 88oakappsUpdate 'sudo docker stats --no-stream | grep honeydue-obs'

# Pod resource usage (k3s side)
kubectl top pods -n honeydue --sort-by=memory

# With stern (if installed: brew install stern)
stern -n honeydue .

# Full state dump for a pod (debugging)
kubectl describe pod -n honeydue <pod> > /tmp/pod-dump.txt
kubectl logs -n honeydue <pod> > /tmp/pod-logs.txt
```

## Future: what to add and when

| Trigger | Add |
|---|---|
| First production incident | Grafana alerting (free, data already there) |
| 10k+ daily users | Loki + Vector for log aggregation |
| Performance incident the histograms can't explain | Wire OTel exporter → Jaeger from the Go app |
| CPU pressure on api pods | Pyroscope continuous profiler |
| Multi-product obs needs | Migrate obs stack to dedicated CX32 ($8/mo) |

The overall philosophy: observability is an investment that compounds.
Add it before you need it, not after. But also don't over-invest at
idle.

## References

- [VictoriaMetrics docs][vm]
- [vmagent kubernetes_sd_configs][vmagent-k8s]
- [Jaeger all-in-one with badger][jaeger]
- [prometheus/client_golang][promclient]
- [Grafana provisioning datasources][gf-prov]
- [Loki][loki] (future)
- [Stern (multi-pod log tail)][stern]

[vm]: https://docs.victoriametrics.com/single-server-victoriametrics/
[vmagent-k8s]: https://docs.victoriametrics.com/vmagent.html#kubernetes-monitoring-with-vmagent
[jaeger]: https://www.jaegertracing.io/docs/1.62/getting-started/#all-in-one
[promclient]: https://pkg.go.dev/github.com/prometheus/client_golang
[gf-prov]: https://grafana.com/docs/grafana/latest/administration/provisioning/#datasources
[loki]: https://grafana.com/oss/loki/
[stern]: https://github.com/stern/stern