I wrote up a post about how we hugely improved the write performance for Bluesky's timelines/following feed. If you want to learn more about how we did it, check it out. Some nuggets in there about embracing imperfection in some parts of a system to scale better. jazco.dev/2025/02/19/i... - ThreadSky

jaz.bsky.social • 100 days ago

I wrote up a post about how we hugely improved the write performance for Bluesky's timelines/following feed.

If you want to learn more about how we did it, check it out.

Some nuggets in there about embracing imperfection in some parts of a system to scale better.

https://jazco.dev/2025/02/19/imperfection/

Comments

jauntywk.bsky.social•100 days ago

It sounds like atm Bluesky is maintaining the timelines for all users all the time.

What if any discussions have happened about letting a user who hasn't logged in a bit's timeline grow stale, backfilling at login time? Pull rather than push?

jauntywk.bsky.social•100 days ago

How far back to timelines go? Is it based on a single factor like time, or number of posts?

jauntywk.bsky.social•100 days ago

(i'd guess count would be better than time, so someone who has a mostly inactive timeline would still have a backlog to look through)

jaz.bsky.social•100 days ago

It's based on # of posts and right now somewhere between 3k-4k posts.

retr0.id•100 days ago

I know Mastodon does this, because if you log into an account for the first time in a while it tells you to wait a bit for your timeline to regenerate

jaz.bsky.social•100 days ago

Yep, I've got a draft proposal for "Timeline Hibernation" that can hibernate a timeline and stop fanning out to it, then regenerate it on-demand once and re-enable fanout to it. Requires a bit of coordination between read and write paths but totally doable and will help reduce costs of our backend.

averagetuna.bsky.social•100 days ago

📌

warhorsefarms.me•100 days ago

@danoleary.bsky.social seems like an article you’d find interesting!

danoleary.me•100 days ago

Thank you! I was give it a read.

michaelakellar.bsky.social•100 days ago

It's enraging that there is plenty of evidence to show that imprefections, flaws, loss, etc. are *needed* for robust systems to flourish; while at the same time we have a Dictator trying to take all of these out of our governmental systems... while watching governmental performance absolutely tank.

michaelakellar.bsky.social•100 days ago

Along with Civics 101, citizens should all be required to take an intro course of thermodynamics and learn about entropy.

bkey1970.keyport.name•100 days ago

... would love to have a writeup on how to get servers fixed in the network, so users could use http://deck.blue - but hey...

owenlacey.dev•100 days ago

Found this really interesting, thanks!

ninji.dog•100 days ago

as a Timeline Enjoyer now I’m worrying that this means I’m missing out on posts because I follow too many people D:

jaz.bsky.social•100 days ago

The threshold is much higher than where you are now, don't worry :)

metaflame.dev•100 days ago

I also feel like this is where feeds based on finer subsets of the social graph (Latest from Mutuals etc.) or e.g. Quiet Posters can do the job well, especially for quieter accounts in the latter case

jaz.bsky.social•100 days ago

Yep, that was my thinking. Someone who follows 20k people is probably not mainlining the following feed but has a bunch of other feeds they use to keep up with the people they care about.

braxuss.eu•98 days ago

Great job. Ideally you would use (to compare with the threshold) a filtered following count, filtering out the quiet users. If someone follows 2000 but 1500 are quiet then they can read the whole timeline. But I guess that's not really necessary. There are low chances of that extreme cases. 😊

leonardopds.com•100 days ago

Very nice, Jaz. I think I have grasped the concept.

Now I wonder how other social medias work. Twitter didn't use to have following limits, so maybe they do something similar? Or just having powerful and numerous servers compensates for slower algorithms and processes? 🤔

cloudhunter.co.uk•100 days ago

Twitter still has following limits! https://help.x.com/en/using-x/x-follow-limit

leonardopds.com•100 days ago

Ops, so I was wrong and X do enforce limits. Nice to know Bluesky didn't go for the easiest path then! Thanks.

drk.wtf•100 days ago

Literally listening to fork around and find out at the point talking about Bsky timeline infra when this post popped up.

mikebeas.com•100 days ago

I understand and appreciate the necessity and simplicity of the approach, but to me it feels like something you should communicate to users who go above that threshold. I would imagine being unaware this was happening and the noticing some posts missing (however unlikely) would raise questions.

lettucelamps.bsky.social•100 days ago

i was also thinking this, but then thought if i were a nefarious user that would be a really valuable number for me. Obviously that number would be scaling with the platform so perhaps not too harmful but yeah it is still public ig, you just have to dig a bit.

elyobo.bsky.social•100 days ago

Hot Shards in Your Area

margotsibley25.bsky.social•100 days ago

Please. How about a site retrieval function on a topical basis. Better than X.

natanael.bsky.social•100 days ago

Do you mean feeds? There are many custom topical feeds here

margotsibley25.bsky.social•100 days ago

Hi, Natanael. I mean say my site. I know I have written in detail about certain subjects. On X I would simply enter say Common Sense in Search and all my past gems would appear. Loved it. Others do feel the same.

jodydonetti.bsky.social•99 days ago

“embracing imperfection”

To accept it and know when to do it (and when absolutely not to) is one of the most undervalued things to learn.

goose.art•100 days ago

cool design, is the "reasonable_follows" number public? i am wondering how lossy my tl is

jaz.bsky.social•100 days ago

rn it's at 4k but it might change based on load, seems to be a decent place for now

goose.art•100 days ago

oh ok cool! i do think this is a great design choice, nobody can realistically tell if they miss something then they follow 10k people lol

goose.art•100 days ago

btw @dame.is might be interesting to integrate this into https://cred.blue as a sort of soft heuristic for overfollowing

dame.is•100 days ago

wait can you tl;dr this for me? you're saying 4K is the point where it's too much? cause i do have a number in @cred.blue's scoring methodology 👀

goose.art•100 days ago

tl:dr, above 4k following people bluesky starts to drop some posts from your tl, at 8k you see 50% of posts at 16k i guess 25%? so basically they say structurally 4k ish is the max reasonable amount to follow

aparker.io•100 days ago

> Hot Shards in Your Area

10/10 heading

moja.blue•100 days ago

The content is fascinating. I'm thinking of translating the blog post into Japanese to share it with Japanese users. What do you think?

jaz.bsky.social•100 days ago

That would be fine :)

moja.blue•100 days ago

thx!

chebykin.org•100 days ago

Amazing article! I've left a comment on reddit, posting here as well: this is an interesting intersection of scaling and ranking! Looking from the ranking perspective, curious how do you decide which post to drop? What if someone post is more important for me to see?

librate.bsky.social•100 days ago

📌

bbl-astrophyscs.bsky.social•100 days ago

📌

jdoda.ca•100 days ago

Interesting stuff. Why this approach instead of a per-user rate-limit on timeline writes (e.g. backed with a token bucket or something)? I mean, my guess would be performance, since this approach just requires reads and maintaining a token bucket requires writes, but are there other considerations?

jaz.bsky.social•100 days ago

Per-user rate limits require much higher synchronization between individual fanout processes than a list of user/following-counts that only covers a tiny percentage of the userbase. The list is maintained every time we process a follow/unfollow in Redis and is refreshed only periodically for reads.

tom.sherman.is•100 days ago

Love this, a blog post that could be used directly as a chapter in a Martin Kleppmann book

Comments

Posting Rules

Reply