A platform is evaluating a service mesh. What should be checked before adopting, and what is the honest justification?

What to check first

1. Is policy drift actually causing incidents? Look at the last ten incidents and count how many a mesh would have prevented. If retry behaviour is inconsistent but nothing has broken, the mesh is solving a theoretical problem.

2. Is the real problem observability? Very often the pain is "we cannot tell which service is slow", and a mesh is being considered because it provides that as a side effect. Distributed tracing, adopted directly, may deliver the actual need at a fraction of the cost — and this is the single most useful question to ask.

3. Could a shared library do it? With few languages, one well-maintained library each is a real option and dramatically cheaper.

4. Can the team operate a control plane? It is a distributed system and a correlated failure domain where a bad push affects everything at once.

5. What is the per-instance overhead across the fleet? Worth calculating rather than assuming.

The honest justification

Mutual TLS everywhere without touching applications, where a compliance requirement demands service-to-service encryption and identity — otherwise a per-language project repeated and maintained forever.

And consistency across languages and teams: retries, timeouts, circuit breaking and outlier ejection implemented once, so a platform-wide decision such as a retry budget becomes possible at all. With library-based policy across many languages, drift is structural rather than a discipline failure.

The deciding factor is usually organisational — the number of teams and languages — rather than technical.

The adoption approach

Incrementally, in value order: mutual TLS and telemetry first, since they deliver most of the benefit with least risk; traffic policy later.

And a specific caution: the control plane's blast radius must be understood before it is trusted. Staged configuration rollout, health-gated promotion on served-traffic signals, and a rollback path that does not depend on the control plane a bad configuration may have broken.