How do you optimize Qdrant for high-throughput production workloads?

Tuning Qdrant for high throughput generally means addressing several independent levers together — indexing parameters, memory layout, sharding, and client-side batching — since no single setting alone typically accounts for the full performance gap between a default setup and a tuned production deployment.

• Tune HNSW parameters: increasing ef (search-time exploration breadth) improves recall at the cost of latency; increasing m (connections per node) at index-build time trades memory and build time for search quality.
• Apply quantization appropriately: reduces both memory footprint and per-query compute cost, especially valuable at large scale.
• Create payload indexes on frequently filtered fields: avoids expensive unindexed scans for common filter patterns.
• Right-size sharding: enough shards to spread load across available nodes, without over-sharding to the point of per-shard overhead outweighing the benefit.
• Batch writes: upload points in reasonably sized batches rather than one at a time, to reduce per-request overhead.
• Use gRPC over REST: lower serialization overhead for high-volume production traffic.

A useful general discipline is benchmarking against representative production-like data and query patterns before and after each change, since the right combination of settings depends heavily on the specific dataset size, embedding dimensionality, and query mix — a configuration tuned for one workload doesn't automatically transfer well to a very differently shaped one.