How Instacart Built a Search for Billions of Products

2025-12-15 · Source: ByteByteGo Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Intermediate, long

Summary

Instacart's search infrastructure evolved from a dual-system approach using Elasticsearch for keyword search and a custom FAISS service for semantic search, to a unified Postgres-based solution. Initially, the company faced challenges with Elasticsearch's denormalized data model, which caused catastrophic indexing loads due to billions of daily writes for rapidly changing grocery item data. They migrated keyword search to Postgres, leveraging GIN indexes and ts_rank, reducing write workload by a factor of ten. Later, they integrated semantic search using the pgvector extension, co-locating vectors with relational data. This consolidation eliminated the need for parallel network calls and application-layer joins, resulting in a 2x speedup, a 6% reduction in zero-result searches, and improved recall by enabling pre-filtering of inventory data before semantic search.

Key takeaway

For AI Architects and Machine Learning Engineers designing search systems with high write volumes and dynamic data, consider consolidating your search infrastructure within a relational database like Postgres using extensions such as pgvector. This approach can significantly reduce operational overhead, improve query latency by bringing compute closer to data, and enhance search quality through advanced pre-filtering capabilities, directly impacting user experience and revenue.

Key insights

Consolidating search compute with data improves performance and simplifies complex, high-write workloads.

Principles

• Match data model assumptions to workload patterns.
• Consolidate compute and data to reduce latency.
• Architectures expire; be ready to adapt.

Method

Instacart migrated from Elasticsearch to Postgres for keyword search, then integrated semantic search via pgvector, enabling real-time attribute filtering and unified data management.

In practice

• Evaluate pgvector for vector search up to 100M vectors.
• Consider Postgres for high-write, relational search needs.
• Prototype new architectures with production-like traffic.

Topics

• Instacart Search Infrastructure
• Elasticsearch Migration
• Postgres Full-Text Search
• Semantic Search
• pgvector

Best for: AI Architect, Machine Learning Engineer, Software Engineer

Related on AIssential

• What (un)exactly do you mean by semantic search? — Specialized vector databases outperform Lucene-based or bolt-on solutions for scalable, semantic search in AI applications.
• The TechBeat: Why UK Retail Real Estate Investors Are Turning to Geomarketing 3.0 (5/18/2026) — Modern tech trends emphasize AI-driven automation, data infrastructure consolidation, and decentralized solutions across various industries.
• Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search — Diverse query initialization significantly enhances agentic search breadth scaling by mitigating first-turn redundancy.
• Larch: Learned Query Optimization for Semantic Predicates — Larch optimizes AI SQL semantic filters by learning evaluation order and predicting selectivities, significantly reducing token costs.
• Announcing Lakebase Search: agent-native retrieval built into Lakebase Postgres — Agent-native search requires instant, real-time retrieval on a single, cost-effective, tiered storage backend.
• What is pgvector? — pgvector integrates vector embeddings and similarity search directly into PostgreSQL, simplifying AI application development.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ByteByteGo Newsletter.