ngit-grasp/docs/explanation/monitoring.md

# Monitoring

ngit-grasp exposes Prometheus metrics at `/metrics` for monitoring WebSocket connections, Git operations, Nostr events, and system health.

## Architecture

```mermaid
flowchart TB
    subgraph ngit-grasp
        HTTP[HTTP Service]
        WS[WebSocket Handler]
        GIT[Git Handlers]
        RELAY[Nostr Relay]
        
        subgraph Metrics Module
            REG[Prometheus Registry]
            CT[ConnectionTracker]
            MC[Metric Counters]
        end
        
        ME[/metrics endpoint]
    end
    
    subgraph External
        PROM[Prometheus Server]
        GRAF[Grafana]
        ADMIN[Admin Browser]
    end
    
    HTTP --> ME
    WS --> CT
    WS --> MC
    GIT --> MC
    RELAY --> MC
    
    CT --> REG
    MC --> REG
    REG --> ME
    
    PROM -->|scrape /metrics| ME
    GRAF -->|query| PROM
    ADMIN -->|view dashboards| GRAF
```

## Configuration

| Option | CLI Flag | Environment Variable | Default | Description |
|--------|----------|---------------------|---------|-------------|
| Metrics enabled | `--metrics-enabled` | `NGIT_METRICS_ENABLED` | `true` | Enable /metrics endpoint |
| Abuse threshold | `--abuse-threshold` | `NGIT_ABUSE_THRESHOLD` | `10` | Max connections per IP before flagging |
| Top N repos | `--top-n-repos` | `NGIT_TOP_N_REPOS` | `10` | Number of top bandwidth repos to track |

## Privacy Model

IP addresses are **never exposed in Prometheus metrics**. The connection tracker maintains per-IP counts internally only for abuse detection:

| Data | Exposed in Metrics? |
|------|---------------------|
| Total connections | ✅ Yes |
| Unique IP count | ✅ Yes |
| Flagged abuser count | ✅ Yes |
| Actual IP addresses | ❌ No (internal only) |
| IP + abuse flag | ⚠️ Logs only (when flagged) |

When an IP exceeds the abuse threshold, a warning is logged but the IP is never exposed via Prometheus.

## Deployment

See [Prometheus Setup Guide](../how-to/prometheus-setup.md) for NixOS configuration and Grafana dashboard provisioning.

## Deletion Lifecycle Metrics

The deletion/recovery operational paths expose these additional metrics:

| Metric | Type | Labels | Description |
|--------|------|--------|-------------|
| `ngit_blacklist_deletions_total` | Counter | `phase`, `result` | Startup blacklist parity deletion attempts/success/failure |
| `ngit_holding_cleanup_runs_total` | Counter | - | Number of holding cleanup passes run |
| `ngit_holding_cleanup_deleted_total` | Counter | `type` | Total deleted objects by cleanup (`metadata`, `payload`, `archive_file`) |
| `ngit_holding_cleanup_last_run_deleted` | Gauge | `type` | Deleted object counts for most recent cleanup pass |
| `ngit_deletion_request_cleanup_runs_total` | Counter | - | Number of deletion-request cleanup passes run |
| `ngit_deletion_request_cleanup_removed_total` | Counter | `type` | Deletion-request payloads and lifecycle metadata removed (`main`, `tombstone`, `metadata`) |
| `ngit_deletion_request_cleanup_outcomes_total` | Counter | `outcome` | Deletion-request cleanup failures and stale/concurrent skips (`failure`, `stale_or_concurrent_skip`) |
| `ngit_recovery_total` | Counter | `result` | Recovery attempts and outcomes (`attempted`, `succeeded`, `failed`, `partial`) |
| `ngit_manual_ejections_total` | Counter | - | Number of operator manual ejection operations |
| `ngit_manual_ejection_deleted_total` | Counter | `type` | Objects removed by manual ejection (`metadata`, `payload`, `archive_file`) |

## Future: Load-Based Sync Scheduling (GRASP-02)

The metrics infrastructure enables future load-based scheduling for GRASP-02 sync jobs:

```mermaid
flowchart TD
    SYNC[Sync Manager] --> CHECK{Check Load}
    CHECK --> MET[Query Metrics]
    MET --> CONN{Connections > N?}
    CONN -->|Yes| DELAY[Delay 5 min]
    CONN -->|No| RUN[Run Sync Job]
    DELAY --> CHECK
```

## Future: Loki for Detailed Logging

For detailed per-repository investigation at scale, consider adding **Loki** (log aggregation):

- Structured logging with tracing crate already in place
- Loki queries enable ad-hoc deep dives (e.g., find all transfers > 10MB)
- Pairs with Prometheus for long-term trends

## Sync Metrics (GRASP-02)

When GRASP-02 proactive sync is implemented, the following metrics will be added to track relay synchronization health. These metrics use in-memory tracking with Prometheus for operator visibility (no database persistence needed for <100 relays).

### Sync Metrics Overview

| Metric | Type | Labels | Description |
|--------|------|--------|-------------|
| `ngit_sync_relay_connected` | Gauge | relay | Connection status (0=disconnected, 1=connecting, 2=syncing, 3=connected, 4=connected_historic_sync_failures) |
| `ngit_sync_connection_attempts_total` | Counter | relay, result | Connection attempt outcomes |
| `ngit_sync_relay_status` | Gauge | relay | Health status (1=healthy, 2=disconnected, 3=degraded, 4=dead, 5=rate_limited) |
| `ngit_sync_relay_failures` | Gauge | relay | Current consecutive failure count |
| `ngit_sync_events_synced_total` | Counter | - | Events synced (newly saved events only) |
| `ngit_sync_relays_tracked_total` | Gauge | - | Total relays discovered |
| `ngit_sync_relays_connected_total` | Gauge | - | Currently connected relay count |
| `ngit_sync_relays_dead_total` | Gauge | - | Relays marked as dead |

### Connection Status Values

The `ngit_sync_relay_connected` metric tracks the connection lifecycle:

- `0` = **Disconnected** - Not currently connected
- `1` = **Connecting** - Connection attempt in progress
- `2` = **Syncing** - Connected, historic sync in progress
- `3` = **Connected** - Connected, historic sync complete, live sync active
- `4` = **ConnectedHistoricSyncFailures** - Connected, historic sync had failures, live sync active, partial data

This allows operators to distinguish between "connected but still catching up" (Syncing) vs "fully synced and live" (Connected) vs "historic sync failures - missing historic data" (ConnectedHistoricSyncFailures).

### Relay Health States

The `ngit_sync_relay_status` metric tracks relay health:

- `1` = **Healthy** - Connected and stable
- `2` = **Disconnected** - Not connected, but no issues detected
- `3` = **Degraded** - Connection problems or unstable after recovery
- `4` = **Dead** - 24h+ of continuous failures
- `5` = **RateLimited** - Rate limit cooldown active (65s)

After a `too many queries` response, `Activated adaptive query-start pacing`
reports the learned per-connection interval. It begins at 600 ms and doubles
only when a distinct later episode proves that pace too fast. Queued starts
unwind during the existing 65-second cooldown before paced recovery;
reconnecting clears that reactive lesson. Independent proactive pacing still
spaces background historic and dependency starts at one per second; persistent
live subscriptions bypass the background gate.
`Falling back to paced REQs for the query-limited connection session` means
NIP-77 remains skipped until reconnect because SDK-managed `NEG-MSG` traffic
cannot be passed individually through the learned gate.

### Example Grafana Queries

```promql
# Relay connection status overview - count by status
sum by (relay) (ngit_sync_relay_connected == 0)  # Disconnected
sum by (relay) (ngit_sync_relay_connected == 1)  # Connecting
sum by (relay) (ngit_sync_relay_connected == 2)  # Syncing
sum by (relay) (ngit_sync_relay_connected == 3)  # Connected
sum by (relay) (ngit_sync_relay_connected == 4)  # ConnectedHistoricSyncFailures

# Relays still syncing (not yet fully caught up)
count(ngit_sync_relay_connected == 2)

# Relays with historic sync failures (missing historic data)
count(ngit_sync_relay_connected == 4)

# Connection success rate over last hour
sum(rate(ngit_sync_connection_attempts_total{result="success"}[1h]))
/ sum(rate(ngit_sync_connection_attempts_total[1h]))

# Event sync rate (newly saved events)
rate(ngit_sync_events_synced_total[5m])

# Relays with high failure counts (potential issues)
topk(10, ngit_sync_relay_failures)

# Relay health overview - count by health state
sum(ngit_sync_relay_status == 1)  # Healthy
sum(ngit_sync_relay_status == 2)  # Disconnected
sum(ngit_sync_relay_status == 3)  # Degraded
sum(ngit_sync_relay_status == 4)  # Dead
sum(ngit_sync_relay_status == 5)  # RateLimited
```

### Example Alerts

```yaml
# Alert if relay stuck in dead state for > 1 day
- alert: SyncRelayDead
  expr: ngit_sync_relay_status == 4  # Dead state
  for: 1d
  labels:
    severity: warning
  annotations:
    summary: "Sync relay {{ $labels.relay }} is dead (24h+ failures)"

# Alert if relay stuck in syncing state for > 1 hour
- alert: SyncRelaySlow
  expr: ngit_sync_relay_connected == 2  # Syncing state
  for: 1h
  labels:
    severity: info
  annotations:
    summary: "Sync relay {{ $labels.relay }} taking >1h to complete historic sync"

# Alert if too many relays are degraded
- alert: SyncManyDegraded
  expr: sum(ngit_sync_relay_status == 3) > 5  # Degraded state
  for: 15m
  labels:
    severity: warning
  annotations:
    summary: "{{ $value }} relays in degraded state"
```

### Design Rationale

**In-memory health tracking with Prometheus visibility** was chosen over database persistence because:

1. **Scale**: <100 relays means per-relay labels have acceptable cardinality
2. **Simplicity**: No database schema, migrations, or cleanup needed
3. **Operator visibility**: Prometheus + Grafana provide better dashboards than custom queries
4. **Restart behavior**: Conservative initial backoff (5s + jitter) avoids thundering herd on restart
5. **Historical data**: Prometheus retains health history; in-memory state only needs current status

See [GRASP-02 Proactive Sync](grasp-02-proactive-sync.md) for full architecture details.

## Rejected Events Index Metrics

The rejected events index tracks rejected repository announcements and state events to prevent wasteful re-fetching during negentropy sync and enable race condition resolution.

### Rejected Events Metrics

All metrics are parameterized by `event_type` label with values "announcement" or "state":

| Metric | Type | Labels | Description |
|--------|------|--------|-------------|
| `ngit_rejected_hot_cache_current` | Gauge | event_type | Current number of entries in hot cache |
| `ngit_rejected_cold_index_current` | Gauge | event_type | Current number of entries in cold index |
| `ngit_rejected_hot_cache_hits` | Counter | event_type | Events retrieved by the explicit invalidation API |
| `ngit_rejected_hot_cache_misses` | Counter | event_type | Explicit invalidations whose full event had already expired |
| `ngit_rejected_hot_cache_expired` | Counter | event_type | Entries cleaned up from hot cache (2 min expiry) |
| `ngit_rejected_cold_index_expired` | Counter | event_type | Entries cleaned up from cold index (7 day expiry) |
| `ngit_rejected_invalidated` | Counter | event_type | Entries removed by the explicit invalidation API |

The invitation dependency-recovery path is deliberately non-destructive and does
not increment the hit, miss, or invalidation counters above. It retains cold IDs
until processing succeeds. Operators can observe exact-ID attempts in structured
logs containing `Fetched purgatory dependencies by exact event ID`; the current
and expiry gauges still cover entries used by both paths.

### Example Grafana Queries

```promql
# Explicit invalidation API hot-cache efficiency
rate(ngit_rejected_hot_cache_hits_total[5m])
/ (rate(ngit_rejected_hot_cache_hits_total[5m]) + rate(ngit_rejected_hot_cache_misses_total[5m]))

# Current rejected events by type
ngit_rejected_hot_cache_current{event_type="announcement"}
ngit_rejected_hot_cache_current{event_type="state"}
ngit_rejected_cold_index_current{event_type="announcement"}
ngit_rejected_cold_index_current{event_type="state"}

# Explicit invalidation activity
rate(ngit_rejected_invalidated_total[5m])

# Explicit invalidation cache hit ratio over time
sum(rate(ngit_rejected_hot_cache_hits_total[5m]))
/ sum(rate(ngit_rejected_hot_cache_hits_total[5m]) + rate(ngit_rejected_hot_cache_misses_total[5m]))
```

### Example Alerts

```yaml
# Alert if cold index growing too large
- alert: RejectedEventsColdIndexSize
  expr: ngit_rejected_cold_index_current > 10000
  for: 1h
  labels:
    severity: info
  annotations:
    summary: "Rejected events cold index has {{ $value }} entries"
    description: "Consider investigating why many events are being rejected"
```

### Two-Tier Architecture

**Hot Cache (2 minutes):**
- Stores full event objects
- Enables immediate re-processing when dependencies arrive
- Cleaned up every 60 seconds
- Memory: ~200 KB typical, ~20 MB worst case

**Cold Index (7 days):**
- Stores metadata only (event ID, pubkey, identifier, reason)
- Prevents re-downloading during negentropy sync
- Supplies exact IDs when dependency-resolvable full events have left the hot cache
- Cleaned up daily
- Memory: ~1 MB typical

### Use Cases

**Race Condition Resolution:**
When a maintainer announcement arrives before the owner announcement:

1. Maintainer event rejected → hot cache + cold index
2. Reciprocal owner announcement enters purgatory → retain the cold ID until recovery succeeds
3. If still in hot cache → immediate policy re-processing
4. If expired from hot cache → exact-ID requests run in parallel across the recursive maintainer relay chain
5. Empty or failed requests keep the ID for a throttled retry
6. Successful processing removes the event from both tiers

**Negentropy Sync Efficiency:**
During sync, cold index IDs are excluded from "missing events" calculation, preventing wasteful re-download of events that will be rejected again.

See [GRASP-02: Integration with Rejected Events Index](grasp-02-proactive-sync.md#integration-with-rejected-events-index)
for the recovery flow.