A platform runs a service mesh. What are the operational realities teams underestimate?

What is underestimated

1. The control plane is a correlated failure domain. A bad configuration push can affect every service simultaneously — a class of incident that did not exist before the mesh. It needs the same staged rollout, health-gated promotion and independent rollback as any other global change mechanism.

2. Upgrades are fleet-wide events. Upgrading the data plane means restarting a sidecar alongside every workload. For a large fleet this is a substantial, risky operation requiring careful sequencing — and it recurs on the project's release cadence rather than on yours.

3. Lifecycle coupling. The sidecar must be ready before the application starts and must drain after it finishes, or requests fail at the edges of every deployment. Getting this wrong produces intermittent errors that are hard to attribute.

4. Debugging moves. "Why did this request fail" becomes a question about proxy configuration and control plane state rather than about application code. The team needs new skills and new tooling, and the transition period is genuinely painful.

5. Resource overhead across the fleet — memory and CPU per instance — which is a real line item at scale and is frequently omitted from the adoption decision.

6. Configuration complexity. Mesh configuration is expressive and therefore easy to get subtly wrong, and the effects can be non-local.

What makes operating it viable

• Adopt incrementally in value order — mutual TLS and telemetry first, which deliver most of the benefit with least risk; traffic policy later.
• Treat control plane changes as production changes, with staging and rollback.
• Version skew tolerance between data plane and control plane, so upgrades need not be simultaneous.
• Escape hatches — the ability to bypass the mesh for a specific workload during an incident.
• Observability of the mesh itself, distinct from the observability it provides.

The honest reassessment

If the driver was observability, a mesh may be the expensive way to get it. Distributed tracing adopted directly frequently delivers the actual need at a fraction of the operational cost — and asking that question before adoption is worth more than any amount of skill afterwards.