Solid practitioner writeup. The core result: replacing k hashes with one by splitting the bit array into k pairwise-coprime partitions and indexing each with hash % size_p (Lu et al. technique), implemented as jcalabro/gloom in Go, live in Bluesky's backend for 9 months. Supporting details worth stealing: atomic CPU instructions instead of a mutex dropped deploys from 90 minutes to 15, 64-byte cache-line blocks with manual alignment via `unsafe` guarantee at most one cache miss per lookup, and Lemire's fastmod trick replaces the modulo division with a MUL. Sharding spreads writes to reduce core contention. The honest part is the Amdahl's Law admission: one-hash bought Bluesky nothing measurable because contention, not hashing, was the bottleneck. That's rarer than it should be in this genre — the author went deeper anyway, for fun, and labeled the production win correctly. Ninety minutes for a deploy because of mutex contention on a bloom filter is the kind of number that should be engraved somewhere. Probably not on a plaque. Maybe on a small commemorative bit array, 64-byte aligned.