A service mesh manages communication between application services, commonly called east-west traffic. Instead of embedding discovery, certificate handling, routing rules, and network telemetry separately into every application, teams configure a shared infrastructure layer. The application still owns business logic, request semantics, and application-level authorization. A mesh is particularly useful when many independently deployed services need consistent communication policies across different languages and runtime environments.
Most meshes separate a control plane from a data plane. The control plane manages configuration, identities, and policies; the data plane handles actual application traffic and enforces those policies. Data-plane components can be sidecar proxies beside individual workloads, shared node-level proxies, gateways, or provider-managed networking infrastructure. A sidecar is therefore one implementation pattern, not a requirement. Applications generally do not send each request through the control plane.
In cloud environments, engineers associate workloads with discoverable service identities and configure how clients reach healthy destinations. For example, an order service can call an inventory service using a stable name while instances are replaced or scaled. Depending on the implementation, routing can distinguish HTTP paths, headers, or service versions. Weighted routing can direct a small proportion of requests to a new version, but safe rollout decisions still require health signals and an explicit rollback process.
Resilience controls can include connection timeouts, request deadlines, bounded retries, load balancing, and circuit breaking. These controls must reflect application behavior. Retrying a payment request can create duplicate charges unless the operation supports idempotency. Retries at multiple layers can multiply traffic during an outage, so configure a coordinated retry budget and keep downstream attempts within the caller's overall deadline. Confirm which controls an implementation actually exposes rather than assuming every mesh has the same feature set.
Security commonly uses workload identities, encrypted connections, and authorization policies that limit which services may communicate. Many meshes support mutual TLS, but transport encryption, peer authentication, and authorization are distinct capabilities. Verify where encryption terminates, how certificates rotate, and whether direct network paths can bypass enforcement. A mesh does not automatically replace network segmentation or application checks such as whether a user may access a particular order.
Observability typically includes request rates, error rates, latency distributions, and traffic logs. These signals help distinguish a slow inventory dependency from a failing order-service instance. Distributed tracing still requires suitable instrumentation and trace-context propagation through applications; proxies alone cannot reliably reconstruct every asynchronous or internal operation. Sampling, log retention, and metric cardinality need explicit limits to control cost.
The trade-offs include additional latency, proxy resource consumption, configuration complexity, and another operational dependency. Test control-plane outages, data-plane upgrades, certificate expiration, and configuration rollback. Some implementations continue forwarding with previously received configuration during a control-plane outage, but discovery updates and credential renewal may eventually fail. For a small application with straightforward communication requirements, ordinary service discovery, load balancing, and application libraries may be simpler than introducing a mesh.