机器学习系统设计
机器学习系统的核心，是训练一个模型来解决商业任务，如预测、分类、排序
建模design：包括优化目标，feature，data，模型结构，评价标准等
系统design：偏重于在线serve模型，包括feature store, ANN, ETL pipeline, MLOps等
1. 面试过程
心态和神态: 自信大方的展现自己的能力，良好的沟通是任何面试都看重的
沟通： 一边白板画框图，一边告知面试官要讲某几个部分。每个部分move前可再次确认
Is there anywhere that you feel I missed?分层思维： 过程中，一层主题讲清楚前，不要陷入任何一部分的细节挖掘。随着问题介绍，data和细节都会明确
深度和广度： 讲到具体部分，尤其是自己熟悉的方面，要主动讲，积极展现自己的知识宽度和深度
trade-off： 不要对需求和场景做主观假设，不熟悉的场景，一定先从头到尾问清楚细节，讲清楚trade-off是能力的重要体现。trade-off可以从从业务(比如预测准确性、长尾预测准确、冷启动效果)和技术角度(scale、latency)出发
2. 回答框架
明确需求 Requirement
functional和non-functional一定确认清楚，否则是明显不合格signal。面试者提的问题往往就能看出水平
场景，功能，目标(engagement or revenue, project goal, project metrics)，约束
scale of the system, user和item有哪些数据和量级
机器学习任务 ML Task
解释如何将需求转化为机器学习问题(如推荐转化为二分类模型和原因)
数据 Data
identify data：training + label, testing + ground truth
获取label: 从交互中收集，人工标注，人工标注加无监督辅助，增强数据
分类任务的positive & negative label
一些可做特征的数据是否有log
数据探讨: bias, 非均衡, label质量
GDPR/privacy: 数据脱敏，数据加密
train/test data和product上distribution不一样怎么办，data distribution随时间改变怎么办
特征 Feature
user, item and cross, context
sparse and dense feature
每个ML组都有不同的embedding set。互相用别人的embedding set，怎么pre-train，fine-train，怎么combine feature非常重要
feature的AB test怎么做？不同traffic
模型 Model
总是从简单baseline开始
每个design的选择，像平时写design doc一样比较不同选项的优劣
模型选择，考虑来自系统的constraint。比如prediction latency，memory。怎么合理的牺牲模型的性能以换取constraint方面的benefit
大多数场景，模型之外都需要额外的策略兜底
评价 Evaluation
offline and online
AB testing
模型的评价，比如：点击，转化，是否有广告？考察的是GMV，还是转化订单？
部署 Deployment
server or device
all users or a part of users
statically, dynamically(server or device) or model streaming
服务 serving
batch prediction or online prediction
监控 monitoring
监控latency，QPS，precision，recall等
Grafana, prometheus
维护 maintain
retrain strategy
全量训练 + 增量训练
3. 面试实例
design a monitoring system to realtime measure ML models, including features, score distribution, qps
abusive user detection
predict app install
业务目标
improve engagement on a feed
improve customer churn
return items from search engine query
cold-start/ position bias/ diversity
multiple task
4. 常见问答
how to scale
Scaling general SW system (distributed servers, load balancer, sharding, replication, caching)
Train data / KB partitioning
Distributed ML
Data parallelism (for training)
Model parallelism (for training, inference)
Distributed training
Asynchronous SGD
Synchronous SGD
Data parallel DT, RPC based DT
Scaling data collection
machine translation for 1000 languages
NLLB
Auto ML (soft: HP tuning, hard: arch search (NAS))
线上线下不一致
不同的数据用什么方式存储
data pipeline怎么设计
deploy
负载均衡和自动伸缩
latency如何优化
这么多server如何deploy，以及如何push新的model version，在更新的时候如何保证qps不degrade
serving
model serving是典型的low latency high qps
Online A/B testing
Based on online metrics we would select a significance level 𝛼 and power threshold 1 – 𝛽
Calculate the required sample size per variation: The required sample size depends on 𝛼, 𝛽, and the MDE Minimum Detectable Effect – the target relative minimum increase over the baseline that should be observed from a test
Randomly assign users into control and treatment groups (discuss with the interviewer whether we will split the candidates on the user level or the request level)
Measure and analyze results using the appropriate test. Also, we should ensure that the model does not have any biases.
If we are serving batch features they have to be handled offline and served at real time so we have to have daily/weekly jobs for generating this data.
If we are serving real time features then they need to be fetched/derived at request time and we need to be aware of scalability or latency issues (load balancing), we may need to create a feature store to lookup features at serve time and maybe some caching depending on the use case.
Where to run inference: if we run the model on the user’s phone/computer then it would use their memory/battery but latency would be quick, on the other hand, if we store the model on our own service we increase latency and privacy concerns but removes the burden of taking up memory and battery on the user’s device.
how often we would retrain the model. Some models need to be retrained every day, some every week and others monthly/yearly. Always discuss the pros and cons of the retraining regime you choose
Monitoring Performance
Latency (P99 latency every X minutes)
Biases and misuses of your model
Performance Drop
Data Drift
concept drift: spam detection
CPU load
Memory Usage
参考
精读
扩展
Last updated