In this session we share our experience of building a real-time data pipelines at Tencent PCG - one that handles 20 trillion daily messages with 700 clusters and 100Gb/s bursting traffic from a single app. We discuss our roadmap of enhancing Kafka to break its limits in terms of scalability, robustness and cost of operation. We first built a proxy layer that aggregates physical clusters in a way agnostic to the clients. While this architecture solves many operational problems, it requires significant development to stay future-proof. With retrospection with our customer and careful study of the ongoing work from the community, we then designed a region federation solution in the broker layer, which allows us to deploy clusters at a much larger scale than previously possible, while at the same time providing better failure recovery and operability. We discuss how we make this development compatible with KIP-500 and KIP-405, and the two KIP (693, 694) that we submitted for discussion.
