Frecency thread feeds (second try) #3798

mxstbr · 2018-08-18T12:18:56Z

Status

WIP
Ready for review
Needs testing

Deploy after merge (delete what needn't be deployed)

mercury
api
hyperion (frontend)

Deploy in that order!!

Related issues (delete if you don't know of any)
Supersedes and thus closes #3727
Supersedes and thus closes #3726
Closes #3718

Based on the conversations in #3727 here is a new approach to frecency thread feeds.

TODO

Add new calculateThreadScore queue
Process jobs in mercury, assigning a score and scoreUpdatedAt for each thread
Add jobs to the queue from mutations changing properties of a thread related to the score
Figure out what to do with all the existing threads
Add sort argument to Community.threadConnection GraphQL type
~~Add sort argument to Channel.threadConnection GraphQL type~~
Look at results with recent db backup and tweak algorithm if necessary
After merging first deploy mercury to production and then the API and the frontend to alpha, then check alpha to see if all looks fine

…24 hours

spectrum-bot · 2018-08-18T12:19:21Z

	Warnings
⚠️	These modified files do not have Flow enabled: mercury/index.js

Generated by 🚫 dangerJS

mxstbr · 2018-08-18T17:26:28Z

These are the results on a not-very-recent prod backup.

Before:

After:

I don't really like these results much yet. The current algo indexes heavily on recent (i.e. last activity), but that means that a thread with 150 participants that's half a year old will shoot up to the top if a single person writes another message. 😕

I think we need to do time boxing when considering the raw content, and count messages that were sent a while ago a lot less than messages that were sent just now. We can do that now! Will dig into that.

mxstbr · 2018-08-20T08:02:38Z

I edited the algorithm to exponentially bias towards recency, based on what Mozilla does. Here's what the results look like on a 15 minute old db backup; first image is sorted by latest, i.e. the old version, second image is trending, i.e. the new version: (note that being ~as good as the old version is good enough for the algorithm since we really just don't want people bumping shitty threads themselves)

Figma

Verdict: Not really a difference in terms of quality at this snapshot in time.

Codepen

Verdict: Algorithmic version seems a little bit better.

React

Verdict: Algorithmic versions looks better.

styled-components

Verdict: Algorithmic versions looks better.

Based on those four hand-picked examples (I didn't check anything else) I say this algorithm looks good for now, let's ship it if the code looks good?

mxstbr · 2018-08-20T08:15:18Z

Just realized my previous migration won't work in production—not sure what to do with all the existing threads just yet...

brianlovin · 2018-08-20T16:29:03Z

api/models/thread.js

+type GetThreadsByChannelPaginationOptions = {
+  first: number,
+  after: number,
+  sort: 'new' | 'trending'


How do we want these options to be defined and expand over time? Is trending the same as popular? Can we ever sort by date? e.g. trending this week/month/year

We'll probably won't be able to do trending this week/month/year with this current setup since we only store the score right now. I think other than that we probably want "TOP", but that's a different calculation altogether.

Not sure, any thoguhts/ideas?

Not sure either, just trying to think ahead a bit for what other constrains might enter the mix. But also don't want to get too caught up to prevent this from moving forward :)

brianlovin · 2018-08-20T16:30:50Z

api/queries/community/threadConnection.js

+export type CommunityThreadConnectionPaginationOptions = {
+  after: string,
+  first: number,
+  sort: 'NEW' | 'TRENDING',


Above we had 'new' and 'trending' but here they are all caps. Should we be consistent with these?

No I don't think so? One is the GraphQL stuff (which has to be all caps) and one is internally the argument for the db query. (which we never do all caps)

Referring to this: https://github.com/withspectrum/spectrum/pull/3798/files#diff-8e46094eed25a21b0755138dc52cf17eR77

No worries either way

brianlovin · 2018-08-20T16:31:00Z

api/queries/community/threadConnection.js

-export default async (root: DBCommunity, args: PaginationOptions, ctx: GraphQLContext) => {
-  const { first = 10, after } = args
+export default async (root: DBCommunity, args: CommunityThreadConnectionPaginationOptions, ctx: GraphQLContext) => {
+  const { first = 10, after, sort = 'NEW' } = args


Ditto to caps vs lowercase

GraphQL stuff

brianlovin · 2018-08-20T16:31:57Z

api/queries/community/threadConnection.js

  const threads = await getThreadsByChannels(channels, {
    first,
    after: lastThreadIndex,
+    sort: sort === 'NEW' ? 'new' : 'trending',


Caps vs lowercase - then we could just say: { sort } if the sort argument is either new or trending

This translates the external GraphQL stuff to the internal db query options.

brianlovin · 2018-08-20T16:32:27Z

api/types/Community.js

  }

+  enum CommunityThreadConnectionSort {
+    NEW


caps vs lowercase?

GraphQL stuff

brianlovin · 2018-08-20T16:32:39Z

api/types/Community.js

    threadConnection(
      first: Int = 10
      after: String
+      sort: CommunityThreadConnectionSort = NEW


caps vs lowercase?

GraphQL stuff

mxstbr · 2018-08-21T13:49:32Z

Oh damn I thought all our GraphQL enums were purely uppercase, that's why I did the external stuff uppercase and internal stuff lowercase—but I just now realized GraphQL enums can also be lower case 🤦‍♂️

Thanks for calling me out on that, you're totally right those should be lower case. On it!

mxstbr · 2018-08-21T13:51:06Z

This should be ready to go, assuming e2e tests pass. The only thing I'm not sure about is how we're going to handle assigning a score to all existing threads—should we just manually add a calculateThreadScore job for every thread in the db?

brianlovin · 2018-08-21T19:24:43Z

should we just manually add a calculateThreadScore job for every thread in the db?

Seems fine to do?

Ship mercury
Run migration
Let mercury process all existing threads
Ship the rest

mxstbr · 2018-08-22T07:37:06Z

Shipping mercury and the the API!

mxstbr added 6 commits August 18, 2018 12:54

WIP: Assign each thread a score

f048d8b

Add calculateThreadScore jobs to the queue from mutations

33e35ca

Add calculate thread score job if threadScoreUPdatedAt is older than …

c47cafc

…24 hours

Store thread.scoreUpdatedAt

34a1443

Fix ms in a day

dc6bd72

Fix flow types

d7beed4

mxstbr mentioned this pull request Aug 18, 2018

Frecency thread feeds #3727

Closed

10 tasks

spectrum-bot bot added the WIP: Building label Aug 18, 2018

mxstbr changed the title ~~Frecency thread feeds #2~~ Frecency thread feeds (second try) Aug 18, 2018

mxstbr added 4 commits August 18, 2018 16:02

Aggressively decay based on lastActive

2578572

Fix typo

1bd214a

Add sort arg to Community.threadConnection

b119225

Run calculateThreadScore once for each thread

cc4c6b4

Rework thread trending score algorithm with exponential explosion

d01c9ae

Remove not-working migration

b073fd6

Top => trending

4f3ed40

spectrum-bot bot added the Ready for Review label Aug 20, 2018

Make sure it is distinct participants

ea79472

brianlovin reviewed Aug 20, 2018

View reviewed changes

threadConnection sort arg lowercase

c3f3d17

Make sure new threads get scores if they dont have any

eea3984

mxstbr merged commit 080caed into alpha Aug 22, 2018

mxstbr deleted the new-frecency-feeds branch August 22, 2018 07:37

Frecency thread feeds (second try) #3798

Frecency thread feeds (second try) #3798

Uh oh!

Conversation

mxstbr commented Aug 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

spectrum-bot bot commented Aug 18, 2018

Uh oh!

mxstbr commented Aug 18, 2018

Uh oh!

mxstbr commented Aug 20, 2018

Uh oh!

mxstbr commented Aug 20, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mxstbr Aug 20, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mxstbr commented Aug 21, 2018

Uh oh!

mxstbr commented Aug 21, 2018

Uh oh!

brianlovin commented Aug 21, 2018

Uh oh!

mxstbr commented Aug 22, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mxstbr commented Aug 18, 2018 •

edited

Loading

mxstbr Aug 20, 2018 •

edited

Loading