You Probably DON’T Need a Service Mesh!
Talkat CloudOpen @ Open Source Summit Europe 2022 · on Thursday, 2022-09-15· at 18:00· Wicklow Hall 2A (Level 2), Convention Centre Dublin, Spencer Dock, N Wall Quay, North Wall Dublin 1, D01 T1·
summary
Argues that Istio-style service mesh (Envoy sidecars per pod) delivers mTLS plus traffic routing features like 95/5 canary splits and fault injection but costs real resources: each sidecar ≈0.35 vCPU and 40 MB memory (so ~35 vCPU and ~4 TB RAM at 1,000 services), control plane ≈1 vCPU and 1.5 GB per replica, and adds ~1.7–2.7 ms latency at p90–p99. Notes boot-order issues (Argo jobs slow/fail), heavy auxiliary tooling, and proposes ambient mesh with a node Z-tunnel for mTLS and namespace proxies for L7 to cut footprint.
AI-written summary · about
chapters (13)
- 0:00Start
- 1:09Introduction
- 3:14Kubernetes
- 3:46Open Source Projects
- 5:39Control Plane
- 10:56Point of a Service Mesh
- 13:34Egress Gateway
- 14:00Traffic Splitting
- 14:17Virtual Service
- 15:46Fault Injection
- 17:18Drawbacks
- 17:45Challenges with How You Design an Architecture Service Mesh
- 22:12Design and Architecture
related talks
34mService Mesh for Developers
38mFewer Moving Parts: Simplify Service Mesh Operations with Istio Ambient Mesh
43mWTF Do We Need a Service Mesh?
32mIstio: Weaving the Service Mesh [I]
09mLightning Talk: Dissecting Service Mesh Overheads
27mService Mess to Service Mesh
28mService Meshes: At What Cost?
01h 03mJust what is a "service mesh", and if I get one will it make everything OK?