Read this interesting retro about discords migration path from Mongo to Cassandra to now ScyllaDB.
https://discord.com/blog/how-discord-stores-trillions-of-messages
@xuu@txt.sour.is I was hoping to read about more technical details of SSTables and how they impact read performance 🤔
@prologic@twtxt.net in the article they say they have a p99 of 15ms reading historical data. Which is pretty nuts.. Aside from having close to 900TB of SSD…
@prologic@twtxt.net in the article they say they have a p99 of 15ms reading historical data. Which is pretty nuts.. Aside from having close to 900TB of SSD…
I suppose the p90/p50 would be higher. Right? I never remember which way that funnel goes.
I suppose the p90/p50 would be higher. Right? I never remember which way that funnel goes.
Pretty sure its the average/p50 that’s lower right? p90/p95 is your end tail. So if most of your latencies are sitting around ~15ms you’d expect your average to be lower, otherwise you’d have a much higher p90/p95 right? 🤔
In any case, yeah that is impressive in and of itself, but I’m not sure what data structures and mechanisms are at play to achieve such impressive read performance? Did I miss it in the article? 🤔