Benchmarking MongoDB Atlas Vector Search for 10M-Vector eKYC Deduplication

CHALLENGE

MCredit needed to evaluate whether vector search could support large-scale 1:N face deduplication for an eKYC workflow. When a new identity is registered, a customer-developed embedding model converts the face image into a 1024-dimensional vector that must be compared against an existing vector repository to identify duplicate identities or potential fraud. Because vector search directly influences the deduplication decision, the architecture had to balance low query latency, sustained throughput, stability and high retrieval accuracy without introducing unacceptable loss from approximation or quantization.

SOLUTION

The POC evaluated MongoDB Atlas Vector Search using dedicated search resources separated from the database workload. The test architecture used MongoDB replica-set nodes for database operations and dedicated search nodes for HNSW-based vector indexing and retrieval, allowing write workloads and vector search workloads to be measured independently. Testing was split into two stages: a 10-million-vector environment focused on latency, throughput, stability and index-build behavior, and a 1-million-vector environment focused on recall, dimensionality and quantization accuracy. HNSW parameters and float32 vectors were evaluated against exact nearest-neighbor search to quantify the trade-off between performance and retrieval quality.

OUTCOMES

• Reached 1,255 QPS at P95 9.85 ms on the 10-million-vector test workload, exceeding the target of at least 100 requests per second below 10 ms P95 latency.
• Completed a 20-hour 16-minute soak test with more than 98 million requests, zero timeouts or errors, and stable search-process memory behavior.
• Achieved Recall@10 of 98.96% with 100 candidates and 99.42% with 200 candidates on the 10-million-vector dataset compared with exact nearest-neighbor search.
• HNSW search completed in approximately 8 ms per query at 10-million-vector scale, compared with approximately 480 ms for exhaustive exact search.
• Testing identified an important production trade-off: float32 vectors preserved the required accuracy, while scalar quantization materially reduced exact-match quality despite lowering index memory requirements.

TECHNOLOGIES

MongoDB Atlas · MongoDB Atlas Vector Search · HNSW · Dedicated Search Nodes · MongoDB Replica Set · float32 Vector Indexing