← Back to Dashboard
1. Telemetry Baselines2. Trace Correlation Across Agent Steps

Telemetry Baselines

📚 Observability and Tracing10 min95 XP

What to Measure on Day One

At minimum, capture request counts, latency percentiles, error rates, guardrail outcomes, token usage, and fallback frequency. Everything here is available from Bedrock's own telemetry surfaces - you just have to turn them on and route them somewhere queryable.

The Baseline Metric Set

MetricWhy it mattersAlarm hint
Invocation countTraffic shape, quota headroomSudden spike = runaway loop or abuse
p50 / p95 / p99 latencyTail latency is what users feelp95 trending toward SLO ceiling
Error + throttle rateCapacity and dependency healthThrottlingException rising = quota pressure
inputTokens / outputTokensThe raw material of your billTokens-per-request drift = context bloat
Guardrail interventionsSafety posture and attack detectionIntervention spike = probing or policy drift
Fallback / escalation rateHow often degraded paths engageRising = upstream quality problem

AWS Building Blocks

  • CloudWatch metrics - Bedrock publishes invocation counts, latency, and token metrics per model out of the box.
  • Model invocation logging - opt-in capture of full request/response bodies to S3 or CloudWatch Logs (mind data classification before enabling in regulated environments).
  • CloudTrail - who invoked what, from which principal - your audit backbone.

Segmentation Is the Difference Between Data and Answers

  • Break down by model, endpoint, and tenant - an aggregate p95 hides one tenant's pathological prompts.
  • Track p50/p95/p99 separately - averages are marketing, tails are engineering.
  • Correlate quality signals with cost over time - cost per successful outcome is the metric leadership actually needs.
Day-one rule: emit a structured log line per invocation - {traceId, model, tokens, latency, guardrailAction, outcome}. Every hard production question later ("why did costs double Tuesday?") is a GROUP BY over this table.
🧪 Knowledge Check
Press 1-4 to select1 of 2
Which latency metric is most useful for user pain?
Average only
p95 and p99 percentiles
No latency metric
Disk throughput