How the feed worked

Two scheduled cloud functions

For every user, calculating word scores
How positive/negative towards what you’re interested in
Taking all contracts, all your views, bets, clicks
- Weighted in interesting ways
- The log of the counts of these
Views are negative, clicks are slightly positive, bets are most positive
Takes words out of contract question, description, tags, creator username
Computes a frequency for each
- TF IDF = “term frequency” * “Inverse document frequency” ⇒ for every word, take the total number of contracts, divide by the number of times where the word appears
  - Scales up unusual words, larger impact on the score
Caches all word vector scores as one document
- Each user has a private cached document. In the private-user there’s a cached collection

Read the word scores, gets the recent fresh data, fetches all contracts in the last 7 days that were active
Dot product of the word score with the frequency of each project
- One subcomponent of the score for the feed ranking
Other component: Activity score
- Last bet and last 3 comments; does the computation based on the last time of the bet and last time of the comment
- Activity score: how much recent, 24h, 7d vol, a bunch of random heurstics
- Comes out with “this is the activity level”
- Multiply activity score by recommendation by the word
How is this implemented to scale?
- Problem: cloud functions have a small amount of time and memory
  - Scheduler: farms out the tasks to another cloud function, batched in terms of 100 users
    - [Austin] Basically mapreduce in cloud functions
  - Will scale - even with thousands more users, going to create a few
Output: “here’s 75 contracts, 3 comments, last bet” ⇒ cached into a single document
- When the user goes home, it fetches single document

Scheduling. Does Vercel do this?
... Not that much reason. Could switch it to Vercel if we want
- When it’s calling each cloud function with each batch of users
- Firebase cloud functions