Statsd Metrics Streaming

/statsd-metrics-streaming

When to Use

• Stream agent performance metrics to Datadog/Graphite without blocking execution
• Track tool call counts, execution durations, error rates per session
• Monitor rate limit hits and sandbox rejections in real time
• Low-overhead UDP fire-and-forget — zero latency impact on agent

Do NOT use for

• Critical delivery metrics (UDP = no acknowledgement, packets may drop)
• High-cardinality tags on every request (DogStatsD tag explosion)

Setup and basic metrics

import StatsD from 'node-statsd'

const stats = new StatsD({
  host:      process.env.STATSD_HOST ?? '127.0.0.1',
  port:      parseInt(process.env.STATSD_PORT ?? '8125'),
  prefix:    'yamtam.',
  errorHandler: (err) => console.error('[statsd] error:', err.message),
})

// Counter: increment on each tool call
stats.increment('tool.calls')
stats.increment('tool.calls', 1, ['tool:fetch', `session:${sessionId}`])

// Gauge: current agent count
stats.gauge('agents.active', activeAgentCount)

// Timer: measure tool execution
const start = Date.now()
await runTool(cmd)
stats.timing('tool.duration_ms', Date.now() - start)

// Set: unique sessions today
stats.set('sessions.unique', sessionId)

Sampling (reduce UDP volume)

// Sample at 10% for high-frequency events
stats.increment('tool.calls.debug', 1, 0.1)
// → sends metric only 10% of the time; StatsD server adjusts rate

DogStatsD tags + histogram

const ddStats = new StatsD({
  host:       '127.0.0.1',
  port:       8125,
  prefix:     'yamtam.',
  global_tags: [`env:${process.env.NODE_ENV}`, `version:${process.env.YAMTAM_VERSION}`],
})

// Histogram (DogStatsD extension)
ddStats.histogram('sandbox.memory_mb', sandboxMemoryMb, ['sandbox:bwrap'])

Wrap tool calls with timing

function measuredTool<T>(name: string, fn: () => Promise<T>): Promise<T> {
  const start = Date.now()
  return fn()
    .then(result => {
      stats.timing(`tool.${name}.duration_ms`, Date.now() - start)
      stats.increment(`tool.${name}.success`)
      return result
    })
    .catch(err => {
      stats.increment(`tool.${name}.error`)
      throw err
    })
}

Anti-Fake-Pass Checklist

❌ No errorHandler → StatsD client throws silently on UDP send failure
❌ Metrics with high-cardinality tags (per-request IDs) → Datadog metric explosion
❌ Missing prefix → metrics mixed with other apps on same StatsD server
❌ UDP drops silently → counter != actual count at high load (use sampling)
❌ stats.close() not called → UDP socket leaks on process exit
❌ timer measured from wrong start point (before async call vs after) → wrong latency