Design a System for E-commerce Search Filtering

To design an e-commerce search system at Amazon scale, I would start by defining non-functional requirements: sub-200ms latency for 99th percentile queries and support for millions of concurrent users. The architecture begins with an ingestion pipeline where product updates are streamed via Kafka to ensure eventual consistency between the transactional database and the search index. For storage, I would deploy Elasticsearch clusters with horizontal sharding based on product IDs to distribute load evenly. Each shard maintains an inverted index mapping terms to document IDs, enabling fast retrieval for text and faceted filters like 'Brand' or 'Price Range'. To handle filtering efficiently, I'd use term vectors for exact matches and range queries for numeric fields. Query routing would involve a gateway service that splits requests into parallel sub-queries across shards, aggregating results locally before returning the top N items. Caching is critical; I'd implement a multi-tier cache with Redis storing frequent filter combinations and result sets to reduce index pressure. Finally, to maintain performance during flash sales, the system must support dynamic scaling of nodes and asynchronous reindexing to prevent blocking writes. This approach ensures high availability and low latency, aligning with Amazon's customer-centric focus on speed and reliability.