跳槽失败，耐心读DDIA

lz硕7年，在五线厂做senior MLE 1年多。工作倒是清闲，但是技术比较落后，大概五个月前开始随缘跳槽。
面了一众公司，包括meta，Uber，Netflix，tiktok等等，0offer。很多都挂在了system design。打算开始潜心读DDIA，修炼内功。
刚读了第一章，做了些笔记，分享出来给需要的人。
大家觉得这种分享有帮助吗？有更好的方式，也请留言告诉我。有用的话，我会继续分享。



第一章
Building Blocks for Data-Intensive Applications
Databases
    Store data so that we can find them later
Caches
    Memorize the result of an expensive operation to speed up reads
Search Indexes
    Allow users to search data by keywords or filter it in various ways
Stream Processing
    Send a message to another process, to be handled asynchronously
Batch Processing
    Periodically crunch a large amount of accumulated data

Reliability
Definition
    Performing functions that users expected
    Tolerating user mistakes
    Good performance under expected load
    Preventing unauthorized access and abuse

Fault vs Failure
    Fault: One component deviating from its specification
    Failure: The whole system stopped

Achieving Reliability
    Think through assumptions and interactions in the system
    Testing at all levels (unit tests, integration tests, manual tests)
    Process isolation
    Allow processes to crash and restart
    Measuring and monitoring system behaviors
    Analyzing system behaviors
    Well-designed abstractions, APIs, and admin interfaces to discourage mistakes
    Decouple places where people make frequent mistakes from ones causing failures
    Allow quick and easy recovery
        Roll back config changes quickly
        Roll out new changes gradually
        Provide tools to recompute data
    Good management practices and training

Scalability
Load Parameters
    Requests per second (RPS) to the server
    Ratio of reads to writes
    Simultaneously active users in the chat room
    Hit rate on the cache
    Distribution of users (e.g., Twitter's top user has 30M followers, causing fan-out for writes)

Describing Performance
    Batch processing: throughput
    Online system: response time
        Average response time
        Response time percentiles (tail latencies are more important for user experience)

Latency vs Response Time
    Response time = service time + network delays + queueing delays
    Latency is the duration that a request is waiting to be handled

Varying Response Time
    Request differences
    Context switch to a background process
    Loss of a network packet
    TCP retransmission
    Garbage collection pause
    Page fault forcing a read from disk
    Mechanical vibrations in the server rack

Calculating Percentiles
    Naive: sort the list of response times every minute
    Forward decay
    T-digest
    HdrHistogram

Achieving Scalability
    Mixture of scaling up (more powerful machine) and scaling out (more machines)
    Elastic system: scale up or down based on traffic
    Factors for system design
        Volume of reads
        Volume of writes
        Volume of data to store
        Complexity of the data
        Response time requirements
        Access pattern

Maintainability
Typical Maintenance Work
    Fixing bugs
    Keeping the system operational
    Investigating failures
    Adapting it to new platforms
    Modifying it for new use cases
    Repaying technical debt
    Adding new features

Improving Maintainability
    Provide visibility into the runtime behavior with good monitoring
    Provide good automation tools
    Avoid dependency on individual machines
    Provide good documentation and operational instructions
    Provide good default behavior but also give admin the freedom to override default values
    Self-healing when necessary but also give admin the freedom to take over
    Exhibit predictable behavior, minimizing surprises

Simplicity
Symptoms of Complexity
    Explosion of the state space
    Tight coupling of modules
    Tangled dependencies
    Inconsistent naming and terminology
    Hacks aimed at solving performance problems
    Special-casing to work around issues elsewhere

Achieving Simplicity
    Simplicity does not mean reducing functionality
    Abstraction: hide implementation details behind a clean facade

Evolvability
Common Changes
    New facts
    Previously unanticipated use cases emerge
    Business priorities change
    Users request new features
    New platforms replace old platforms
    Legal requirements change
    Growth of the system

最近也面了一定量大公司的MLE/SDE in ML相关的岗位，全部铩羽而归，感觉也都是挂在了SD（有些明确说的，有些是自己觉得的，因为coding都能完整做出来），不过还是有些心得可以分享。


首先同意上面说的，DDIA是要读的。读了九死一生，不读十死不生。然后MLE或者搞ML Data Infra相关的，推荐各位读一读下面三本书：
Machine Learning Engineering (link)
Designing Machine Learning Systems: An Iterative Process for Production-Ready Applications (link)
Machine Learning Design Patterns: Solutions to Common Challenges in Data Preparation, Model Building, and MLOps (link)

多说一句Alex Xu的书 (link)，当字典看很好，面试直接背答案你就死了，因为根本来不及展开，而SD最基本的要求是要complete high level design。

并且可以多搜搜看YT上一些system design interview的视频。不是说那些视频讲得都是好的，相反把看视频当作自己是在参加一场别人的design review，看完后可以想想这个视频讲得有什么问题，怎么改进。当然有些讲的真的不错的可以好好学习。

以及如果各位在大厂或者技术不错的中厂的话，内部的各种design doc也是准备SD很好的资料。有机会参加各种design review也很有帮助，因为SD模拟的就是个design review的过程。

大家加油，面试尤其是SD这种偏主观的面试，水平到了一定程度的，能不能成其实都看面试官的一念之间。很多时候你挂了不是技术不到位，只是你答的不是面试官想要的。毕竟45分钟-1个小时，设计的东西肯定不可能面面俱到的。所以面试的时候可以多套套话，看看面试官到底想要一个足够复杂的high level design，还是深入某些components，还是想多聊聊model。何况不少来面SD的面试官自己其实火候也完全不到位，自己也不理解自己出的这题，只是拿一个现成的design来看你能不能对上，所以除非面试者是该系统设计者本人，不然怎么答都是错的。这种情况下，失败了也不必气馁。概率问题而已，不是我们能控制的，继续努力便是。

这句话不仅是告诉给诸位的，也是告诉给我自己的。

补充内容 (2024-04-29 05:10 +08:00):
强调一下，我这里的MLE指的是做system和MLOps的，不是主要做模型的。做模型的在我的context里被称为Scientist

读的时候要多多上网看看有没有更新更好的技术，引自作者的BLOG:

```
My book continues to sell well, with now over 230,000 copies sold, and reviews continue to be very positive. However, it is gradually showing its age – it was published in 2017, but I wrote the first few chapters around 2014/15, so they are now almost a decade old. Moreover, I have learnt a lot in the meantime, and there are quite a few things in the book that I would now say differently
```

https://martin.kleppmann.com/2024/01/04/year-in-review.html

MLE为什么读ddia？又不是后端工程师

senior MLe不读一读Alex的一本新的ML interview么

你们别学了，卷不动了

读ddia是有用的，但是现在ddia已经远远不够了。

MLE senior 也要面system design吗？

MLE真正要面sd的多吗？

匿名用户 发表于 2024-04-01 19:10:21
读ddia是有用的，但是现在ddia已经远远不够了。

那现在要怎么准备啊？都考察什么现在？

匿名用户 发表于 2024-04-01 19:36:30
那现在要怎么准备啊？都考察什么现在？

插个眼，我也想知道现在sd要怎么准备