We're building a document retrieval system and evaluating Qdrant vs Pinecone. The open-source angle is appealing but I'm curious about real-world scaling—how many vectors are you storing and what's latency like under load? Also, does anyone have experience hosting it on their own infrastructure vs managed solutions?
I'd recommend testing Qdrant locally first with your actual document volume—I've seen teams hit solid sub-100ms latency with millions of vectors on modest hardware, but it really depends on your embedding dimensions and query complexity. Self-hosted gives you cost savings at scale but requires DevOps overhead; if you want to benchmark performance, consider loading sample data through mcp-server-data-exploration to simulate real retrieval patterns before committing. What's your expected QPS and vector dimensionality?