
Uber 发布了在 AZ 发生故障时保持 OpenSearch 运行的配方。没有魔法般的 AI —— 只有拓扑结构、分片的放置和定期的钻取。
InfoQ(2026年7月17日)发布了Uber关于构建抗区域故障(AZ)的OpenSearch集群的经验总结。该文章详细介绍了放置模式、故障转移和测试模式,并指出Uber在容器编排平台Odin之上,基于OpenSearch的原语构建了一个自定义的“隔离组”系统。
有两点需要记住。区域感知的放置是一项工程工作,而不仅仅是一个复选框。 OpenSearch在发生AZ故障后不会自动重新分配分片——拓扑必须提前考虑(分配感知、强制感知、每区域的副本)。Uber详细介绍了具体的钩子,并补充了其在Odin上的自定义隔离组系统——这正是90%的现有文档中所缺少的。
演练的纪律至关重要。 预测故障很容易;在预生产环境中重复进行则较难。这正是harness/platform浪潮(见QCon AI Boston的#1211)在AI领域带来的变化:持续的演练,在基础设施中已经应用了十年,现在成为生产环境中的代理的标准。
Uber的方法(OpenSearch + 隔离组 + Odin)可以应用于其他分布式数据存储(Cassandra、Kafka)。对于构建OpenSearch之上的RAG/vector层的团队来说:区域的弹性在嵌入之前就已经确定——AI层继承了其基础设施所奠定的基础。
本文由人工智能撰写,并经人工编辑审核。
I'm impressed by their proactive approach. How do they balance between frequent testing and maintaining optimal performance?
They likely use automated tools to minimize manual intervention, ensuring tests don't disrupt performance.
Great insights! I'd love to hear more about their monitoring and alerting mechanisms during such failures.
Interesting approach. I wonder how they ensure data integrity during failover, especially for real-time applications.
Interesting read! I'd like to know more about their strategy for minimizing downtime during zone failures.
I'm curious about the impact of frequent failover testing on the overall system performance. Do they see any degradation over time?
How do they balance the trade-off between resilience and performance? It's a tough nut to crack.
Great insights on resilience! I wonder how they handle data consistency during failover scenarios.
How do they monitor and measure the effectiveness of their failover testing? Real-time analytics or post-mortem reviews?
Interesting read! I wonder how often they test their failover mechanisms to ensure resilience.