知识卡片
延迟预算决定微批或常驻
实时摄取的常驻计算会持续产生成本。若业务能容忍分钟级延迟,可采用微批降低空转开销;需要亚秒级响应时再评估流处理。选型先给出可测的端到端延迟目标,而不是把“实时”当作默认要求。
这张卡片看不懂?
这张卡片的反馈已满 10 条,暂时无法继续提交。
已有反馈 · 0 / 10 条
还没有人提交反馈。
原书出处
参考来源
- 位置:《Practical Lakehouse Architecture》第7章《The Big Picture: Designing and Implementing a Lakehouse Platform》(源文件:_epub-src/OEBPS/ch07.html)
- 结论依据:原文对照常驻实时计算的成本、分钟级微批及低延迟流处理选项。
- 原始内容:Real-time ingestion needs an “always on” compute cluster, which can result in higher costs. If latency is not a concern, try executing these workloads as micro-batches (one to five minutes) to lower the cost. Spark Structured Streaming is a good option for real-time ingestion. However, if you need an extremely low-latency solution (sub-seconds), you can consider