Bit of Systems @bitofsys

Joined October 2026

Tweets

74
Followers

2
Following

110
Likes

21

Bit of Systems @bitofsys

15 hours ago

> the fixes - LRU-K / 2Q: a key has to be seen twice before it counts as hot - segmented LRU: new keys start in a probation segment - LFU: rank by how often, not how recently Redis has allkeys-lfu for exactly this case.

0 0 0 2 0

View Details

Bit of Systems @bitofsys

15 hours ago

> the weakness: scan pollution One big scan (a backup, a report, a SELECT *) reads thousands of keys once. Each one jumps to the head. Your hot keys slide to the tail and get evicted. After the scan, the cache is full of junk you won't read again.

1 0 0 4 0

View Details

Bit of Systems @bitofsys

15 hours ago

> your cache is full > a new key shows up > something has to go > which one do you throw out? LRU's answer is simple. Kick out the key nobody touched for the longest time. Here's how it works and why it's O(1)

1 0 0 7 0

View Details

Bit of Systems @bitofsys

2 days ago

you don't need a new database to shard. pick a few keys, hash them, fake it with views first, keep rollback cheap. great write-up by Sammy Steele on the Figma blog: figma.com/blog/how-figma…

0 0 0 7 0

View Details

Bit of Systems @bitofsys

2 days ago

> did it work? about nine months to shard the first table. live in September 2023. the cutover cost 10 seconds of partial availability on primaries. none on replicas. no latency or availability regressions after.

1 0 0 5 0

View Details

Bit of Systems @bitofsys

2 days ago

> Figma's tables hit billions of rows > They'd maxed out vertical scaling and splitting tables across databases > One table no longer fits on one machine, and you can't make the machine any bigger > What do we even do?