Course contentsShow
← System Design
Lesson 865 of 1,91922. Service MeshFree lesson

Performance Overhead: Latency and Throughput

Quantify the performance cost of service mesh sidecars, typically adding 1-5ms latency and 10-20% CPU overhead per request.

Performance Overhead: Latency and Throughput

You'll learn: How to quantify the tangible performance costs that service mesh sidecars introduce into your request path.

The Cost of Convenience

Service meshes provide powerful features—mTLS, observability, traffic management—but these capabilities don't come free. Every request passing through your mesh must traverse the sidecar proxy sitting next to each service, adding measurable overhead to both latency and resource consumption.

Typical Performance Impact

When you deploy a service mesh, expect these performance characteristics:

Latency overhead: Each proxy hop typically adds 1–5 milliseconds to request processing time. For a single request traveling from Service A to Service B, you're adding two proxy traversals (egress from A's sidecar, ingress to B's sidecar).

CPU overhead: The sidecar proxy consumes resources for TLS encryption/decryption, routing decisions, and telemetry collection—typically 10–20% CPU overhead per request compared to direct service-to-service calls.

Memory footprint: Each sidecar process requires its own memory allocation, usually 50–100MB per pod at minimum.

Why This Matters

Think of the service mesh like adding security checkpoints at every border crossing. Each checkpoint makes travel safer and more controlled, but inevitably slows the journey. For high-throughput systems processing thousands of requests per second, these milliseconds compound quickly.

A service chain spanning five microservices might add 10–25ms total latency just from mesh overhead. For latency-sensitive applications (real-time bidding, gaming, financial trading), this overhead can be unacceptable.

Planning for Overhead

Always benchmark your specific workload with and without the mesh. The actual overhead depends on proxy configuration, TLS cipher selection, and telemetry verbosity. Some teams find 2ms acceptable; others cannot tolerate it.

Key Takeaway: Service mesh sidecars typically add 1–5ms latency and consume 10–20% additional CPU per request—measure this overhead against your performance budget before adopting a mesh.