知识卡片
标签基数限制指标细分
状态码、有限操作种类等标签能把请求量分解成可比较序列;延迟或无限用户ID作为标签会制造大量序列。时延应使用数值分布,个体诊断用日志或追踪保存标识。Prometheus每组标签组合都是一个时间序列,因此关联ID不能无界地加入普通指标标签。
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。
在整书笔记中的位置
原书出处 · 1 处
- Chapter 14. Monitoring and Observability Patterns · 查看原书依据
Labels need to be a limited set of discrete values, while latency is a value that can take on many different arbitrary values. Request latency is a separate metric, and it is generally modeled as a histogram. Using the histogram metric for request latency, you can understand the
参考来源
- 位置:《Designing Distributed Systems, 2nd Edition》第14章《Chapter 14. Monitoring and Observability Patterns》(来源项目 book-sources;源文件:extracted/Designing Distributed Systems, 2nd Edition/_epub-src/OEBPS/ch14.html)
- 结论依据:原文已指出时延和客户ID不宜作标签;以官方Prometheus规则消解后文“每个指标加correlation ID”的冲突。
- 原始内容:Labels need to be a limited set of discrete values, while latency is a value that can take on many different arbitrary values. Request latency is a separate metric, and it is generally modeled as a histogram. Using the histogram metric for request latency, you can understand the