Pinterest Reduces Costs by Up to 30% with SSD Serving and Quantization in Manas Embedding Search Platform

The Pinterest engineering team shared the evolution of the embedding search stack for their proprietary distributed search platform, Manas. The existing HNSW-based architecture required loading the entire index into RAM, causing costs to surge proportionally with corpus size. To address this, Pinterest introduced three key technologies: Quantization, SSD-based Serving, and Multi-embedding Retrieval.

Maximizing Memory Efficiency

Quantization, which compresses vectors into low-bit integers, was applied to all major use cases, reducing embedding index memory by over 50% and serving infrastructure costs by 20–30%. Benchmark results on a 100-million-item GraphSage dataset showed that while Scalar Quantization (SQ) offers a lower compression ratio than Product Quantization (PQ), it maintains high Recall above 90%, providing an excellent balance. Additionally, the implementation of Linear Scaling SQ eliminated decoding bottlenecks, further reducing per-query computing resources by 10–15%.

SSD-Based Serving Architecture

By introducing a method to serve indexes directly from NVMe SSDs instead of RAM, memory usage was reduced by 10 times and CPU costs by 40%. In particular, the optimization technique combining PQ with the SPANN algorithm is 4.5 times faster than standard SPANN, with QPS more than 3 times higher and latency one-third lower compared to DiskANN. In pre-evaluations on over 5 billion Pin recommendation data points, CPU time was reduced by more than 40% compared to HNSW, while keeping Recall loss below 5%.

Introduction of Multi-embedding Retrieval

To overcome the limitations of single-vector representations in existing Two-tower models, Late Interaction models (such as ColBERT) were introduced. Documents and queries are represented as vector lists, and 'Sum of MaxSim' logic is applied to support rich token-level similarity scoring. The Manas serving stack has integrated these multi-embeddings as a new query type and is currently validating them through pilot use cases. Future plans include actively adopting SPANN, ColBERT, and GPU-based search systems to evolve into a scalable and cost-efficient architecture.