Pinterest的分析作为Druid的一个平台（第2部分，共3部分

Jian Wang, Jiaqi Gu, Yi Yang, Isabel Tallam, Lakshmi Narayana Namala, Kapil Bajaj | Real Time Analytics Team

王健，顾佳琪，杨毅，Isabel Tallam，Lakshmi Narayana Namala，Kapil Bajaj | 实时分析团队

This series is three parts. To read part 1, click here. Part 3 coming soon.

这个系列有三个部分。要阅读第一部分，请点击这里。第三部分即将推出。

In this blog post series, we’ll discuss Pinterest’s Analytics as a Platform on Druid and share some learnings on using Druid. This is the second of the blog post series, and will discuss learnings on optimizing Druid for batch use cases.

在这个博文系列中，我们将讨论Pinterest在Druid上的分析平台，并分享使用Druid的一些心得。这是该系列博文的第二篇，将讨论关于优化Druid的批量使用案例的心得。

Learnings on Optimizing Druid for Batch Use Cases

为批量使用案例优化Druid的经验之谈

During the process of onboarding different use cases, we found many critical system metrics were missing in Druid. In response, we added metrics on usage of processing threads, merge buffers, rows in memory, etc., to help us get a clearer idea of the bottlenecks affecting capacity provisioning and to identify opportunities to make improvements.

在加入不同用例的过程中，我们发现Druid中缺少许多关键的系统指标。作为回应，我们增加了关于处理线程、合并缓冲区、内存中的行等使用情况的指标，以帮助我们更清楚地了解影响容量配置的瓶颈，并确定改进的机会。

Tiering Based on Request Pattern

基于请求模式的分层

Druid supports loading segments into different server pools based on segment timestamps, and higher end hosts can be used to serve more frequently accessed segments. Initially, we didn’t have quantitative analysis on the request time range distributions other than heuristics provided by clients, so the tiering was somewhat arbitrary. Later we logged all Druid requests to a data source to analyze, and it has helped us make smarter decisions on tiering. Below is an example of one of the use cases that we used to load all six months’ segments into memory optimized hosts. After query pattern analysis, we found 98% of the requests hit the most recent 35 days, so we shifted segments older than 35 days ago to io optimized-based hosts without impacting SLA while saving infrastructure cost.

Druid支持根据段的时间戳将段加载到不同的服务器池中，更高端的主机可以用来提供更频繁的访问段。最初，除了客户提供的启发式方法外，我们没有对请求时间范围分布的定量分析...