Why Service Mesh

The Problem Evolution

 stage 1 — few services:
   HTTP calls with a shared client library. done.

 stage 2 — the estate grows (polyglot, many teams):
   every cross-cutting need multiplies by LANGUAGE COUNT:

   need              java-lib  go-lib  py-lib  node-lib...
   retries/timeouts     ✓        ✓       ?        ?
   mTLS                 ✓        ?       ✗        ?
   metrics/traces       ✓        ✓      half      ?
   circuit breaking     ✓       roll    ✗        ?

   drift is inevitable: python's retry policy ≠ go's.
   security fixes = N dependency upgrades × M services.

The Mesh Answer

 move ALL service-to-service plumbing OUT of application
 code into a dedicated infrastructure layer:

 [app]──localhost──►[proxy sidecar]══mTLS══►[peer proxy]──►[app]
                          ▲
                   [control plane] pushes config/policy/certs

 capabilities become FLEET-WIDE POLICY, not per-language code:

 □ mutual TLS + identity everywhere (zero app changes)
 □ uniform retries/timeouts/breaking as declarative config
 □ L7 routing: canaries, mirroring, header-based splits
 □ consistent golden telemetry from every hop

 write once in the mesh; all languages inherit.

The Trade You’re Making

 GAINS: consistency, security defaults, traffic control,
        polyglot peace — real and compounding with scale.

 COSTS: proxies on every pod (CPU/RAM), an extra hop,
        a control plane to operate, NEW failure/debug modes,
        organizational learning curve.

 the honest decision rule:
   <10-20 homogeneous services → libraries are cheaper;
   polyglot estates with compliance needs → mesh earns it.
   adopting because "Netflix/K8s does" = ceremony tax.

What Mesh Is NOT

 ✗ a business-logic tier (services stay owners of logic)
 ✗ an API gateway replacement (north-south vs east-west —
   complementary edges of the traffic picture)
 ✗ magic observability (it gives STRUCTURE; you still design
   dashboards/alerts)
 ✗ automatic resilience correctness — bad timeout policy in
   YAML fails exactly like bad code did

Interview Framing

“Do we need a service mesh?” scored shape: name the library-drift problem as THE motivation, draw data/control planes briefly, present gains AND costs symmetrically, give the size-based decision rule, distinguish east-west from north-south. This question grades whether you adopt infrastructure from PROPERTIES or FASHION — show the arithmetic either way.