Pinterest's Manas Platform Evolves from HNSW to Quantized SPANN

Pinterest Engineering reworked its Manas distributed search platform, serving tens of billions of embeddings across 80 clusters. SQ and PQ quantization cut HNSW indexes by 59–74%, while SPANN with SSD-based posting lists delivered 3x the QPS of DiskANN and saved 20–30% in serving costs.

• PQ compressed HNSW indexes by 74%, IVF by 93% at 70–80% recall
• SQ reduced HNSW by 59%, IVF by 75% with recall above 90%
• SPANN with PQ gave 3x DiskANN QPS at 1/3 latency, 5% recall drop
• On 5 billion embeddings, SPANN saved over 40% CPU time