felgueres / x_docs

Refs
Ads API. https://documenter.getpostman.com/view/7379786/SzRw1WTU#6a2980d8-e027-4643-93f3-de790c76413f Twitter Arch: https://miro.com/app/board/uXjVPBnTJmM=/ Twitter Eng: https://blog.twitter.com/engineering Dev docs: https://developer.twitter.com/en/docs/twitter-ads-api How twitter works. https://twitter.com/justinhendrix/status/1594031025658675202

Functionality

• campaign management
• custom audiences
• creatives
• advertising analytics

Factors for ads pricing

• auction
• ad score
• bidding

Terminology

• price per billable action is not fixed, it's variable
• billable action: action that you pay for in your campaign

Ad score = quality + bid, this is metric that determines which ad will be served

Bidding campaings

• Campaign objectives:
  • followers: more people see ad
  • engagements: more likes, retweets, replies and clicks
  • website traffic: more visitors to website
  • keywords: users with high intent (beta!) - likely current work
• Bidding options:
  • automatic - auto optimizer
  • maximum bid - exactly how much to pay for a BA
  • target bid - average daily cost
• Quality score:
  • ad is given quality score determined by resonance, relevancy and recency
  • strong copy and visuals are extremely important
  • resonance: engagement
  • relevance: alignment with audience
  • recency: up-to-date content
• Other factors influencing pay per billable action:
  • Size of target audience
  • How many other advertisers are competing

Targeting affects audience size

• additive - "OR": Interest, username, movies, events, keywords, topics
• restrictive - "AND" (makes pool smaller): geolocation, age, render, device, platform, language

How it works

• If you win the auction you get to make impressions
• When user does a billable action, you pay with a second price auction model You only pay whats needed for your ad to beat the 2nd place ad in the auction eg. if you bid $1.4 and beat an otherwise identical ad bidding $1.0, your campaign is charged 1.01 on the billable action

How to build it

• Distributed key-value store
• Costs are in terms of communication and data shuffling
• Complexity varies on dataset size, number of nodes and distribution of data across nodes

Practice

• Submit ad, think improvements in process

• Build a 2nd price auction system

  • Considerations to scale it
    • Data struct to keep it things sorted? Maybe a balanced binary tree to access at in log n
    • Lazy sort: sort only when
  • What hapens when someone withdraws or expires bid? How does this affect the data struct used?

• What does the API look like? [x] What are the core endpoints? [x] What does the design look like?

• Requirements gathering Scale and performance

  • Bids per second or per auction
  • Concurrent auctions and bidders Data consistency
  • Immediate consistency needed or eventual consistency acceptable?
  • Consistency requirements? No bids lost or incorrectly ordered? Reliability
  • How would you handle failures and network partitions?
  • System uptime and fault tolerance? Monitoring and maintenance
  • What logging and reporting is needed?

Twitter
Post Tweet Use publish new message to followers (4.6K requests/s avg, 12K request/sec at peak)
This is likely closer to 6K requests/s in 2024
Write ops
Home Timeline
User can view tweets posted by people they follow (300K requests/s)
Mostly read ops
Home Timeline Tweets + Ads
User can view tweets posted by people they follow AND targeted ads

How to create an ad

1. Choose objective
2. Create an ad
3. Customize delivery

Problems Ad relevance and personalization. How to use user data to predict ads to show? Input: Feature engineering on user profile Posts Interactions Followees Times when most active

Compute: Match ads vs. profile Topic modeling -> keywords of interest, themes (best for coarse categories of interest) Recency modeling -> Grab latest 5-10 to have recency bias -> best for time sensitive products Match ads vs. user's profile with embeddings Predict engagement metric for given content Predict is_user_monetizable

Output: Relevant Ad inventory

Timeline Mixer. Updates user-specific cache Input: User profile Followee new tweets Ad inventory Compute: Rank Output: List of tweets + relevant ads

Distributed system to serve billions under 500ms. Caching

Where does AI fit in? Creatives Targeting train model to predict interests, categories

wishlist for X human curation works, let me train my feed modes: entertainment, learning, work ability to ask questions about X content transcripts on video and spaces summarize my timeline X creator studio: help promote my work why things trend

Questions to ask? GenAI fits into creatives? Whats the limiter?