Twitter, MAU = 300M, DAU = 150M

Scenario

Concurrent User:

• QPS = DAU * average request per user / total seconds per day = 150M * 60 / 100k = 90k
• Peak QPS = 3 * QPS = 270k
• Read QPS = 300k
• Write QPS = 5k

Service

• split into micro-services

• define features for each service

e.g.

• user service
  • login
  • register
• tweet service
  • post tweet
  • news feed timeline
• media service
  • upload image
  • upload video
• friendship service
  • follow
  • unfollow

Storage

• SQL / NOSQL / File System?

• schema, STAR schema, (de-)normalize

e.g.

• user service: SQL (id, username, email, password)
• tweet service: NoSQL (id, userId, content, createdAt)
• media service: File System
• friendship service: SQL / NoSQL (from_userId, to_userId)

Scale

Optimize

• pull / push? normalize / de-normalize
• more features?
• special case, such as, Lady Gaga, Inactive users...

Maintenance

• handle failure
• scalability

News Feed

• Twitter / Facebook / RSS reader / Wechat Friend Circle...

pull model

• read news feed
  • get followings (friendship table)
  • get tweets from followings (tweets table)
  • merge N sorted array
  • analysis: N DB reads + merge k
• post a tweet
  • 1 DB write (tweets table)
• cons:
  • N * DB.getTweets(following, 100) // N DB read is expensive and blocking process
  • how to??? -> cache
    • cache each user's timeline (top 100 tweets) (reduce pull time)
    • cache each user's news_feed (reduce merge time)
• pros:
  • 1 DB write

push model

• news feed table
  • Id, ownerId, tweetId, createdAt
• read news feed
  • 1 DB read from news feed table, top k latest
• post a tweet
  • insert to tweet table
  • get followers from friendship table
  • insert and fan out to followers in news feed table
• cons:
  • when followers is huge, Lady Gaga, push may take very long
  • how to??? ->
    • rank followers by weight (e.g. last login time)
    • mark as "star" user. "star" user do not push tweets to news feed table. Followers pull from "star" user timeline
      • merge results from news feed table and "star" user timeline
• pros:
  • get followers and fan out can be async
  • 1 DB read

什么时候用Push?

• 资源少

• 想偷懒，少写代码

• 实时性要求不高

• 用户发帖比较少

• 双向好友关系，没有明星问题(比如朋友圈)

什么时候用Pull ?

• 资源充足

• 实时性要求高

• 用户发帖很多

• 单向好友关系，有明星问题

Follow / Unfollow

• Follow一个用户之后，异步地将他的Timeline合并到你的News Feed中

  • Merge timeline into news feed asynchronously.

• Unfollow一个用户之后，异步地将他发的Tweets从你的News Feed中移除

  • Pick out tweets from news feed asynchronously.

Store Likes

• Tweet table
  • id, userId, content, createdAt, likeNums, commentNums,, retweetNums