System design Trending:

Failure oriented,

Tolerate errors,

Anytime kill itself and will not impact others, micro architecture

How

The process of defining the architecture, component, modules, interfaces, and data for a system to satisfy specified requirements

General Process

Conceptual Desing, Transparentcy black, Marco level

Logical desing, Transparentcy grey,

Physical design, Transaprentcy White, Micro Level

5 Steps for cracking a desin(SNAKE)

based on example of Netflix

• Scenario: ask features / interfaces / DAU or QPS
  • Enumerate functions: Register/login, Play movie, Movie recommendation
  • Sort top functions: Play movie(get channels, movies in channel, play a movie in a channel)
  • providing API: method name
• Necessary: constrain / hypothesis
  • Ask for daily active users
  • Predict Users
    • average concurrent users = daily active users/ daily seconds * average online time
    • Predict peak users = 6 * average concurrent users
    • Predict peak users in 3 months = Predict Peak user * 2
  • Predict traffic(band use)
    • traffic per user = 3mbps
    • Max peak traffic = predict peak users * traffic per user
  • Predict Memory
    • Memory per user: 10KB
    • Max daily memory = daily active users * memory per user
  • Predict Stroge
    • Total movie: 10000
    • Average movie size: 90min -> 120min * size per minute
    • Movie strage = Total moview * average movie size
• Application: split application / service / module / algorithm
  • Replay the case, add a service for each request
  • Merge the services
• Kilobit: data
  • Append dataset for each request below a service
  • Choose storage types: MySQL, MongoDB, Files?
• Evolve (improve or scale): sharding, optimize, special case
  • Analyze (consider some)
    • with Better: constrains
    • with Broader: new cases
    • with Deeper: details
    • from Permance
    • from Scalability
    • from robustness (reliability)
  • Go back by evolving accordingly

Compare Pull/Push Model

Use Push Model

• less resource
• less coding
• low requirement for realtime
• less user post
• Bi-direction follower/following relationship, without super star (similar to broadcast)

Use Pull Model

• more resource
• realtime requirement
• a lot of user post
• single direction follower/following, has super star(similar to broadcast)

Pull Model: (need sync DB read)

Every time request news feed, server query and merge all related data then return

Get News Feed => n DB read, sync have to wait

Post Feed => 1 DB write

Process:

• Request me news feed
• Get followings
• Get news feed from every followings (DB sync read, user must be waited)
• Merge and return

Optimize:

• cache users news feeds, 1000 feeds, n DB sync read -> n cache read
  • trade off: not cache all
• 100 = Memcached QPS / MySQL QPS

Push Model:

server side add a table for every user, including all related data, once request news feed, just fetch and return existed data

struct NewsFeedTable {
    var id: Int
    var userId: Foreign Key
    var feedId: Foreign Key
}

Get News Feed => 1 DB read

Post News Feed => n DB write, n = followings, async write

Process:

• Post a feed
• Server insert feed into DB
• Server async send feed to its followers
  • Get its followings
  • Fanout: insert feed to its followers

Optimize

• Disk is cheap, storing NewsFeedTable
• Inactive User waste storage,
  • sort by last login time,
  • ignoring some inactive user when insert feeds
• Super star(ton of followers) is slow to push,
  • let fans of super star use pull model
  • others use push model
  • Trade off: How to determine Super star?
    • <1m followers: push
    • 1m - 10m: pull + push (user merge when pull)
    • >10m: pull

Design New Feed Summary

• Requirement Analysis: Scenario
• Data/Scale Prediction: Needs
• Application Graph
• Schema Design
• Push Model vs Pull Model
  • Super star problem
  • Inactive users
• following & unfollowing
• Normalize & De-normalize
  • count Likes: de-normalize, count in feed model, every day do a calibration
  • comments: Normalize, count in DB
• Hot Spot
  • How about cache dropping hot spot item?, leasing get