gRPC is a high-performance framework using Protobuf for efficiency, but it has limitations in browser support and tooling maturity compared to REST.
gRPC supports various streaming patterns over persistent HTTP/2 connections, which can cause traffic hotspots with traditional Layer 4 load balancing.
Choose client-side load balancing for low-latency internal services and infrastructure-based load balancing for external APIs that require a clear demarcation point.
A review of basic, load-based, and hash-based algorithms reveals that options like "least outstanding requests" can outperform simple round robin for uneven loads.
Long-lived streaming connections can prevent traffic from being distributed to newly scaled instances, making traditional CPU-based autoscaling ineffective.
Use tools like grpcurl for functional API testing with proto files and ghz for comprehensive performance and load testing of your gRPC services.
A practical example shows how separating unary and streaming gRPC calls into different Kubernetes services and target groups solves uneven load distribution.
Successfully load balance gRPC by being mindful of long-lived sessions, understanding client traffic patterns, and selecting L7-based algorithms when possible.
27:17Bastian Eicher
28:16Alexis Yushin
30:26Mario-Leander Reimer
32:19Kevin Klues
20:29Simon Auer
20:39Duan Lightfoot
28:03Federico Fregosi
28:55Ingo Weichsel
Jobs that call for the skills explored in this talk.
Wolters Kluwer
Alphen aan den Rijn, Netherlands
Remote World
Grafana Labs
Redruth, United Kingdom
BridgePhase, LLC
Arlington, United States of America
Grafana Labs
GULP Information Services GmbH
Köln, Germany
