DATACORE SEARCH PLATFORM

Vector Search Indexer

Custom implementation of HNSW indexing for sub-10ms similarity search at scale.

P99 LATENCY

8ms

QPS

10,000

INDEX UPDATE

<2s

COST REDUCTION

60%

CHALLENGE

Existing vector search relied on a managed service with unpredictable latency spikes during index rebuilds. Query P99 exceeded 200ms under concurrent write load, breaking real-time recommendation SLAs.

SOLUTION

Built a custom HNSW indexer in Rust with memory-mapped segment files and incremental index updates. Write path batches embeddings into append-only segments; read path merges segment graphs lazily with copy-on-write semantics.

RESULT

P99 query latency stabilized at 8ms under 10k QPS with concurrent writes. Index rebuild time reduced from 45 minutes to incremental updates under 2 seconds. Infrastructure cost dropped 60% versus managed alternative.

STACK

RustHNSWmmapTokioPostgreSQL
AJ

Distributed Systems & AI Infrastructure

© 2026 ALIREZA JAFARTASHBUILT WITH PRECISION · REV. 2026.06