Zalando

Search Context

Zalando is Europe’s largest fashion platform. Their Search & Browse team powers:

Catalog search (full-text + filter) for millions of daily requests
Zalando Assistant — an AI-powered product recommendation system that queries the same search infrastructure

This coupling means search degradation has compound blast radius: not just UX failures, but AI assistant failures and dark campaigns.

Facet-heavy catalog: fashion filtering (brand, size, color, price, fit) generates aggregation-heavy queries that have different load profiles from result retrieval
Cache stampede risk: layered caches across all tiers create expiry coordination problems under load
Dual consumers: search serves both direct user queries and AI assistant product fetches — blast radius spans both
Scale: millions of daily queries with latency SLAs in the tens of milliseconds

Elasticsearch as the core retrieval engine (Base Search layer)
Four-layer architecture: Base Search → NER Query Builder → Catalog API → Search API + Algorithm Gateway
Each layer independently cacheable; ES coordinator nodes provide additional caching
NER for entity recognition → implicit filter injection