Trace Context Propagation

What Must Travel

 context = small metadata riding WITH every request:

 W3C traceparent format (the standard):
   traceparent: 00-<trace-id>-<parent-span-id>-<flags>
   version-  traceid           spanid        sampled-flag

 also common: b3 single/multi-header (zipkin heritage),
 vendor formats historically — w3c is the convergence point.

 propagation = read incoming → create child span → write
 outgoing to every downstream call. EVERY call. one missed
 branch = severed subtree in your waterfall.

The Propagation Matrix

 where context must survive, and how:

 HTTP/REST:      headers (middleware does it — verify!)
 gRPC:           metadata; interceptors standard
 MESSAGE QUEUES: message headers/properties — THE gap:
     producer injects into kafka headers / sqs message attrs
     consumer extracts BEFORE processing starts
 ASYNC frameworks: task/thread-pool hops must CAPTURE and
     RESTORE context (executor wrapping!) or spans orphan
 SCHEDULED jobs: synthesize root spans per execution
 EXTERNAL calls out: propagate INTO third parties? only if
     they support it; else new root with correlation linkage

 the thread-pool case is the sneakiest:
 async servlet/futures losing context across executor
 boundaries — spans appear parentless, waterfalls fragment.
 solved by context-aware executors; discovered by missing-
 span dashboards.

Baggage: Context Beyond Tracing

 baggage propagates arbitrary KEY-VALUES alongside:

   baggage: tenant=acme, tier=premium, debug-session=x9

 uses that earn their keep:
 □ tenant/tier available DEEP in call chains without
   parameter-drilling through every signature
 □ routing/policy decisions at mesh level per baggage
 □ forced-sampling flags for special debug sessions ✓

 discipline (baggage abuse is real):
 □ tiny payloads only (every hop carries it)
 □ SECURITY review of what's included (it leaks into logs/spans)
 □ documented schema; not a free-form dumping ground

Verification Culture

 broken propagation is silent until you look:

 □ CONNECTED-RATIO metric: % of traces with >1 service span;
   drops after deploys flag severed instrumentation ✓
 □ ORPHAN-span alerts: roots appearing mid-flow where children
   were expected
 □ PROPAGATION TESTS in CI: mock multi-hop flows asserting
   context continuity through every transport the service uses
 □ NEW-TRANSPORT CHECKLIST: any new client/queue/library gets
   propagation wired + tested as part of adoption, not later

 treat context like authentication middleware:
 infrastructure-owned, tested, assumed-present everywhere.

Interview Framing

“Traces show disconnected fragments between our API and worker services” scored shape: context anatomy explained (traceparent), queue-header injection/extraction identified as the likely break, thread-pool-executor subtlety named, connected-ratio monitoring proposed, baggage mentioned scoped-with-discipline. Propagation questions test whether tracing was WIRED or WISHED — the verification culture answer separates them.