Advanced1 min readUpdated August 1, 2026
What breaks first when you go from a million vectors to a billion?
Stub — heading outline only, full article to follow.
Memory is usually the first wall
Index build time stops being a background task
Tail latency degrades before average latency does
Sharding strategies once a single node can't hold the index
What actually changes in your operational playbook at this scale
Related questions
Quantizing embeddings: how much can you compress before quality drops?
Product and scalar quantization can shrink vector memory footprint by 4–32x, but recall degrades gradually then falls off a cliff past a certain compression ratio.
Why can't you just use Postgres for vector search?
You often can, with pgvector — but it stops scaling gracefully well before purpose-built vector databases do, and knowing where that line is matters.