Mesh Timeouts

Timeouts Everywhere by Default

 mesh gives every hop an EXPLICIT timeout — the "no call
 without a limit" rule enforced fleet-wide:

 route-level:
   timeout: 3s          ← whole-request envelope at THIS hop
   retries.perTryTimeout: 1s  ← each attempt's slice

 defaults matter: many meshes ship conservative defaults
 (istio's historic 15s-ish) that silently shape your SLOs.
 AUDIT them like code, don't inherit them blindly.

The Budget Chain in a Mesh World

 same arithmetic as always, now distributed across proxies:

 user budget 1000ms
   gateway→orders:    timeout 500ms (retries: 2×200ms)
   orders→payments:   timeout 300ms
   payments→fraud:    timeout 100ms ← tightest leaf

 rules re-verified per hop:
 □ parent > child sums (or parent cancels children)
 □ per-try ≤ overall with room for actual retries
 □ DB/cache calls inside services still need THEIR limits —
   mesh covers service-to-service; app-side calls remain yours

 DEADLINE PROPAGATION GAP to know about:
 meshes don't automatically shrink downstream timeouts by
 elapsed upstream time unless headers carry deadlines and
 both ends honor them. check YOUR stack; assume gaps otherwise.

Timeout Failure Modes Specific to Meshes

 □ DOUBLE-WRAP confusion: app sets 5s client timeout AND
   mesh route says 15s → effective 5s; dashboards show
   "mesh timeouts" never firing → false confidence.
   ONE owner per limit layer; document who owns which.

 □ TIMEOUT vs RETRY interaction misconfigured:
   timeout 2s, retry ×3, per-try unset → one slow attempt
   eats everything; retries never happen (see retries lesson)

 □ LONG-POLL/streaming endpoints killed by blanket L7
   timeouts → exempt via explicit long-timeout routes or
   L4 pass-through for streams. enumerate these endpoints!

 □ IDLE timeouts on connections (TCP level) vs REQUEST
   timeouts conflated in reviews — different dials!

Endpoint class	Timeout posture
Interactive reads	tight (p99×margin)
Checkout writes	bounded + idempotent retries
Reports/exports	async-ify; no giant sync timeouts
Websockets/streams	idle-based, exempt from request caps

Interview Framing

“Random 504s from the mesh after our p99 got worse” scored shape: budget-chain audit approach (parent/child arithmetic), double-wrap ownership check first, retry/per-try interaction verified second, streaming-endpoint exemptions third. Mesh-timeout incidents are almost always CONFIG INTERACTION bugs — demonstrating the audit ladder is the competence signal.