Read on Twitter
The image below is the spec sheet circulating for Parler's entire hosting needs (via @th3j35t3r).
As a cloud technologist who builds big distributed systems, some remarks on the infrastructure footprint:
As a cloud technologist who builds big distributed systems, some remarks on the infrastructure footprint:
First, let's just start with some back-of-the-envelope math.
Parler has ~8M users. Let's say every single person posts an average of once per day.
(In practice, if it's like other networks, most people post nothing, some people post a little, and a small number post a lot.)
Parler has ~8M users. Let's say every single person posts an average of once per day.
(In practice, if it's like other networks, most people post nothing, some people post a little, and a small number post a lot.)
Let's say each and every post is a 200 KB image with some text.
That's 8M Ă— (say) 200 KB/post Ă— 1 post/24 hr / 60 min/hr, or a bandwidth of about 10 Gb/minute.
That's 8M Ă— (say) 200 KB/post Ă— 1 post/24 hr / 60 min/hr, or a bandwidth of about 10 Gb/minute.
Parler says their internal traffic is ~400 Gb/minute, which is absolutely wild. Now, certainly some of that is for things like internal systems, observability metrics, and so on, but it's not gonna be 40Ă— the core traffic of the app.
This is a wildly bloated network footprint.
This is a wildly bloated network footprint.
How about the infrastructure footprint? They're running 40 instances with 512 GB RAM each, for a total of 20 TB of RAM.
As a ballpark, if you devoted all of that to just caching feeds, that's enough RAM to give 2 MB to every single Parler user and cache their feed in memory.
As a ballpark, if you devoted all of that to just caching feeds, that's enough RAM to give 2 MB to every single Parler user and cache their feed in memory.
Overall, the bloat is off the scales here.
Parler seems to have undertaken aggressive vertical scaling in the early days instead of making their systems more efficient, then turned to horizontal scaling when they ran out of headroom.
Parler seems to have undertaken aggressive vertical scaling in the early days instead of making their systems more efficient, then turned to horizontal scaling when they ran out of headroom.
All in all, I'd be very surprised if Parler js back online anytime soon, unless they figure out a way to make their application much leaner and cut the bloat (or if these numbers aren't accurate).
Thanks for reading.
Thanks for reading.
