Leveraging Service Meshes for Accelerating Serverless Workflows
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Thursday, 2020-11-19· at 17:40· Intrado Virtual Event Platform,·
summary
Claims that co-located functions cut workflow latency are backed by measurements: direct calls between services hit a median of 620 µs (1 KB), Knative Serving about ~2 ms (1 KB), Eventing sequences ~1.45 ms and Serving+Eventing ~4.59 ms median (1 KB), while large payloads rise to ~90 ms median (2 MB) for the combined pattern versus ~40–54 ms for simpler paths. Implementation uses Knative, Istio/Envoy and a prototype called knx micro functions that packs multiple functions per container and leverages the Envoy sidecar for locality-aware dynamic rebalancing.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:29Serverless Workflows
- 2:55Platform design
- 3:29Basic Services
- 4:14Knative Serving
- 5:16Knative Eventing Sequence
- 6:44Eventing + Serving
- 7:22Different data sizes
- 8:16Summary
- 8:55Outline
- 9:30Overheads in transferring large data
- 12:17Need for internal load rebalance
- 13:57Load rebalancing with data transfer
- 15:26Dynamically configurable communication
- 20:13KNIX Microfunctions (knix.io) • Colocate functions of a workflow inside single container wrapped in a native service Provides custom local message bus within the container
- 21:14Source code
related talks
21mServerless computing using Knative Open Source
21mWhat is the Future of Serverless Computing?
40mThe State of Serverless Computing
30mPerformance Tales of Serverless
27mServerless Performance on a Budget
27mServerless on Kubernetes: WASM, Knative, or Regular Autoscaling – What Actually Works?
34mKubeFlow’s Serverless Component: 10x Faster, a 1/10 of the Effort
35mStop Calling Knative Serverless!