Mesh Observability

The Uniform Telemetry Bonus

 every proxy emits consistent metrics per request:

 PER SERVICE-PAIR automatically:
 □ envoy_requests_total by response-code
 □ request duration histogram (p50/p90/p99...)
 □ bytes sent/received
 □ connection counts, TLS handshake stats

 because EVERY hop passes a proxy:
 - service-to-service latency MAP emerges without app changes
 - success rates per edge (who breaks when?) instantly visible
 - version-level splits (canary gates read these!)

 [gateway]─99.9%─►[orders]─99.2%─►[payments]─100%─►[bank]
        the graph of health, measured uniformly ✓

Tracing Through the Mesh

 proxies propagate trace CONTEXT (b3/w3c headers) hop-to-hop
 and emit spans:

 [span: gateway]→[span: orders]→[span: payments]

 what mesh tracing gives FREE:
 - context propagation correctness across languages ✓
   (the eternal manual-instrumentation pain)
 - proxy-side spans showing MESH-ATTRIBUTED time

 what it does NOT give free:
 ✗ application-internal spans (your DB queries, business steps)
 ✗ trace BACKENDS — mesh exports; you still run jaeger/
   tempo-class storage + choose sampling strategy
 
 complete tracing = mesh propagation + APP-level spans.
 meshes start the firewood; apps add the flame.

What the Mesh Cannot See

 honest blind spots:

 □ INSIDE-process behavior: thread pools, GC pauses, queue
   depths within services — app metrics remain mandatory
 □ PAYLOAD semantics: which business operation failed?
   (mesh sees /orders POST 500s; not WHICH order or why)
 □ DATABASE/cache hops: app→db traffic bypasses mesh L7 view
   unless db sidecars exist (rare) — instrument drivers
 □ QUEUE/event flows: kafka-class traffic isn't HTTP;
   separate monitoring discipline applies

 rule: mesh telemetry = the INTER-SERVICE skeleton.
 attach organ-level app metrics or you'll know WHERE the
 body hurts but never WHY.

Building on Mesh Metrics

Use	Mesh metric source
SLO dashboards per service	request-rate/error/duration series
Canary gate automation	per-subset delta comparisons
Dependency maps	service-pair edges auto-discovered
Capacity planning	per-edge RPS/bytes trends
Incident triage	”which edge degraded FIRST?” timelines

 cardinality warning: per-route/per-version labels multiply
 series fast. trim default label sets aggressively; keep
 high-cardinality exploration in traces, not metric labels.

Interview Framing

“Prove observability improves after mesh adoption” scored shape: uniform-metrics-per-edge story with dependency-map payoff, tracing-propagation-vs-spans distinction stated precisely, blind-spot list volunteered (app internals/payload semantics/db), cardinality-trimming note. The graded skill is knowing exactly where mesh telemetry ENDS — overclaiming is the tell of non-operators.