Performance Optimization - Maximum Search Speed for 55M Vector Collection #7717
Replies: 1 comment
|
For 55M vectors, I would avoid tuning only from the client thread count. First separate three bottlenecks: HNSW search cost, payload/filter IO, and cluster fan-out/reduce overhead. Things I would check in order:
If you need 99% recall/precision, the query-time
You have
Track p50/p95 latency and CPU utilization per node. The best RPS point is often not the largest batch.
With
75 segments for 55M points is not obviously wrong, but search can improve after optimization reduces avoidable segment fan-out. Make sure there are no pending optimizer tasks and compare performance before/after a controlled optimize cycle in a staging copy.
You are using A useful benchmark matrix would be: |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hello everyone!!!
Could you please advise how to achieve maximum search speed for a 55 million vector collection while maintaining 99% precision? We have a 10-node cluster with 265GiB RAM per node.
Current performance: approximately 30 RPS with 16 concurrent threads.
Collection Configuration:
Search Implementation:
Vector search is implemented via batch search through API with batches sent across 16 separate threads:
endpoint = f"collections/{collection_name}/points/query/batch"
What optimizations would you recommend to improve search performance from the current 30 RPS?
Thank you!
All reactions