qdrant-indexing-performance-optimization

Speed up Qdrant uploads by using batch upserts of 64-256 points per request and 2-4 parallel upload streams. Server-side, increase shard count (3-12) for independent update workers and create payload indexes before HNSW builds.

Qdrant does not build HNSW indexes immediately; small segments use brute-force search until they exceed indexing_threshold_kb (default 20 MB). This slower search window is by design, not a bug, and resolves once the optimizer builds the index.

Reduce HNSW build time by lowering m (default 16 is sufficient for most cases), reducing ef_construct to 100-200, and keeping max_indexing_threads proportional to CPU cores. GPU indexing is also supported for faster builds.

No, setting m=0 is legacy and can drop an existing HNSW index, causing long reindexing. Instead, set indexing_threshold_kb very high during bulk load and restore it afterward so the optimizer builds the index once loading completes.

For multi-tenant collections split by a payload field like tenant_id, use payload_m to build HNSW indexes only per tenant subset instead of a global index. Skipping the global HNSW index significantly reduces total indexing time.

Large merges and HNSW rebuilds legitimately take hours on big datasets. Check progress via the optimizations endpoint (v1.17+), monitor CPU and disk I/O since HDDs are not viable, and inspect logs if optimizer_status shows errors like disk full.