Metrics

Metrics are the numeric time series behind your product: request rates, token throughput, latency histograms, custom counters, and events. They are OpenTelemetry- and Prometheus-compatible and served by Hanzo O11y, tenant-scoped by org.

Emitting metrics

Push metrics through the OTLP endpoint on the gateway, or expose a Prometheus /metrics endpoint and let O11y scrape it. Either way the same hk-* key authenticates and X-Org-Id scopes the series.

from prometheus_client import Counter, Histogram

inference_requests = Counter(
    "inference_requests_total", "Total inference requests", ["model", "status"]
)
inference_latency = Histogram(
    "inference_duration_seconds", "Inference latency", ["model"]
)

with inference_latency.labels(model="qwen3-4b").time():
    result = run_inference(prompt)
inference_requests.labels(model="qwen3-4b", status="ok").inc()

For OTLP, point your exporter at the gateway — metrics land at POST /v1/metrics:

export OTEL_EXPORTER_OTLP_ENDPOINT=https://api.hanzo.ai/v1
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer hk-...,X-Org-Id=org_123"

Querying with PromQL

Query stored series with PromQL over REST:

curl "https://api.hanzo.ai/v1/metrics/query" \
  -H "Authorization: Bearer hk-..." \
  --data-urlencode 'query=rate(inference_requests_total[5m])'

Common expressions:

# 99th percentile latency
histogram_quantile(0.99, rate(inference_duration_seconds_bucket[5m]))

# Error-rate percentage
100 * rate(inference_requests_total{status="error"}[5m])
    / rate(inference_requests_total[5m])

Events and sessions

Alongside numeric series, O11y rolls up product events and per-user sessions so you can chart adoption next to reliability. User-level rollups (trace volume, tokens, and cost per userId) are available at /v1/o11y/users, and multi-turn runs are grouped as sessions.

Related

• Logs and Traces — the other two observability signals
• Sessions — multi-turn runs derived from traces
• O11y — dashboards, alerting, and SLOs over these metrics