Roblox’s detailed post‑mortem reveals that a three‑day outage was caused by a Consul streaming bug and a design flaw in BoltDB’s freelist, which together created CPU contention and latency spikes on its massive on‑premises infrastructure, leading the team to disable streaming, add a second data‑center, and redesign their architecture.
This article introduces service governance fundamentals, explains how etcd’s Raft‑based consensus and BoltDB storage work, compares etcd with ZooKeeper and Consul, and describes Baidu’s large‑scale, high‑availability, high‑performance service‑governance platform built on these technologies.
This article explores service governance fundamentals, examines why etcd’s Raft‑based consensus and BoltDB storage make it ideal for large‑scale systems, compares it with ZooKeeper and Consul, and shares Baidu’s practical architecture, performance tricks, and operational metrics for high‑availability, high‑performance service management.
This article reviews the evolution of etcd’s read‑write mechanisms, explains the “Fully Concurrent Read” feature introduced in etcd 3.4, and presents experimental results showing how it dramatically reduces expensive read latency and improves overall throughput in Kubernetes clusters.
This article explains the inner workings of etcd, covering its Raft consensus implementation, node election and state management, as well as the persistent storage layer built on BoltDB, and includes detailed code excerpts to illustrate each component.
