A platform moves cross-cutting concerns into a sidecar process alongside every service. What does this solve, and what new problems appear?

What it solves

Cross-cutting concerns implemented once rather than per language. In a polyglot estate, retries, timeouts, circuit breaking, service discovery, mutual TLS and telemetry implemented as libraries means one implementation per language — and therefore inconsistent behaviour, drift, and upgrade cycles measured in quarters.

Moving them out of the process gives:

• One implementation for every language, so policy is uniform by construction.
• Policy changes without application redeploys, pushed through a control plane.
• Consistent, comparable telemetry at every hop — frequently the most valuable outcome, because it makes latency attribution across a large service graph possible for the first time.
• Uniform identity and mutual TLS, which is otherwise a per-language security project.

The underlying problem being solved is organisational: library consistency across many teams and languages is unattainable by agreement, so it is made structural.

What new problems appear

1. An extra network hop each way, adding latency and CPU per instance. Small per request and meaningful at fleet scale.

2. Resource overhead per instance, which is a real line item when multiplied across a large fleet.

3. A control plane to operate — itself a distributed system, and a correlated failure domain: a bad configuration push can affect every service simultaneously. This is a new class of incident that did not previously exist.

4. Debugging moves from reading application code to reasoning about proxy configuration, which is a genuine learning curve and a source of confusing incidents.

5. Lifecycle coupling. The sidecar must start before the application is ready and stop after it drains, or requests fail at the edges of a deployment.

When it is not justified

Twenty services in one or two languages. A shared library is simpler, cheaper, has no extra hop, and is easier to debug.

The sidecar earns its cost when the number of languages and teams makes library consistency impossible — which is a people problem solved with technology, and worth being honest that this is the actual justification rather than a technical one.

The adoption sequence that works

Mutual TLS and telemetry first, which deliver most of the value with the least risk, and traffic policy later once the team is comfortable operating the control plane.