High‑Conversion Overseas E‑Commerce Search Using Alibaba OpenSearch

Project Background and Value

With data volumes growing exponentially and search becoming a core infrastructure for digital transformation, GoTerra – a leading Southeast Asian e‑commerce platform – faced challenges in multilingual semantic understanding, heterogeneous data indexing, ultra‑low latency across public networks, and elastic cost control.

Client Pain Points

Multilingual and semantic gap: Complex morphology, spelling variations, and localized semantics made out‑of‑the‑box search engines insufficient, requiring costly custom tokenizers and semantic recall engines.

Unified indexing of heterogeneous data: Structured attributes, unstructured text, and cross‑modal vectors could not be efficiently managed by a single relational or inverted index while preserving strong consistency.

Public‑network ultra‑low latency: P99 latency needed to stay below 70 ms despite cross‑region calls, demanding high internal computation efficiency and cache strategies.

Elastic cost control: GoTerra needed resource usage to match traffic peaks and valleys precisely, avoiding fixed‑cost over‑provisioning and preventing accidental service interruptions.

Overall Solution Architecture

A four‑layer, stream‑batch integrated architecture was designed to decouple computation, indexing, and service layers, achieving optimal performance‑cost balance.

Data Ingestion Layer – Stream & Batch Dual Channel: Real‑time field updates (seconds) and offline batch updates (T+1) are handled via upsert commands and MaxCompute Schema, providing unified data pipelines.

Index Construction Layer – Multi‑Modal Index Strategy: Custom Southeast Asian multilingual analyzer and N‑gram fallback improve recall; vector indexing uses white‑box HNSW or quantization clustering; OBJECT/NESTED types support complex structures, enabling field‑type‑based index splitting.

Search Service Layer – Parallel Multi‑Path Recall & Extreme Optimization: Simultaneous keyword, dense vector, and filter queries are executed; kernel‑level optimizations such as inverted index memory locking and filter/query rewriting ensure low latency and high throughput even over public networks.

Operations Layer – Elastic Scaling & Security Loop: Time‑based elastic scaling automatically adjusts LCU quotas to match traffic curves; instance‑level deletion protection and full‑link monitoring create a robust production‑grade operation shield.

Key Feature Implementations

4.1 Multilingual Analyzer & N‑gram

Problem: Southeast Asian languages have complex affixes, abbreviations, and spelling variants that degrade recall.

Solution: A locale‑aware tokenizer precisely extracts terms and proper nouns; an N‑gram analyzer handles out‑of‑vocabulary and misspelled words, building a tolerant inverted index.

Result: Search accuracy and recall improved dramatically, eliminating >60 person‑days of custom tokenizer development.

4.2 White‑Box Vector Indexing

Problem: Need semantic similarity matching with transparent control over indexing and retrieval.

Solution: Users can select cosine similarity, configure HNSW or quantization clustering, and define namespace partitions for logical isolation. Fine‑grained parameters such as TopK and 距离阈值 are exposed for quality tuning.

Result: Eliminated black‑box risk, allowing algorithm engineers to map business semantics directly to index strategies.

4.3 OBJECT/NESTED Composite Types

Problem: Complex nested fields (e.g., product attributes, user profiles) cause mismatched queries if modeled incorrectly.

Decision Path:

OBJECT: Flattened storage for fields that do not need independent querying, boosting query performance.

NESTED: Preserves internal relationships for independent queries on array elements, preventing cross‑object result pollution.

Result: Accurate modeling avoided result mismatches and enabled efficient management of complex data.

4.4 Multi‑Path Parallel Search

Problem: Single‑path recall cannot satisfy both exact keyword matching and semantic exploration.

Solution: Concurrent execution of keyword, vector, and attribute filters; results are normalized and fused according to configurable priority, recall count, and weight.

Result: Combined “precision” of keyword recall with “breadth” of vector recall, delivering a massive uplift in search experience.

4.5 Offline Data Processing

Problem: Vector generation and large‑scale historical data reconstruction must be decoupled from online queries.

Features:

High‑throughput upsert batch operations simplify offline sync.

Seamless MaxCompute Schema integration automates ETL from data warehouse to search engine, supporting T+1 full vector rebuild without impacting online services.

Result: Built an elastic, low‑maintenance algorithm pipeline.

Extreme Performance Optimization

Initial P99 latency was ~250 ms. A three‑layer systematic optimization (resource → query → index) reduced OpenSearch‑side P99 latency to <70 ms, maintaining stable user experience despite public‑network fluctuations.

Cost Optimization & Operational Safety

6.1 Time‑Based Elastic Scaling

Business traffic exhibits tidal peaks; fixed high‑peak resources cause waste. A time‑dimension elastic policy automatically shrinks LCU quotas during low‑traffic periods and expands before peaks, aligning cost with usage.

6.2 Instance Deletion Protection

To prevent accidental instance deletion, a logical forced protection (mirroring Alibaba Cloud SLB) activates when only one online version remains, and a second‑factor confirmation is required for any delete operation.

Standardized Migration Path (Best‑Practice Framework)

Requirement Analysis & Gap Assessment: Technical review of functional needs, performance KPIs (QPS, P99), ES baseline, data model, and query patterns.

Feature Validation & POC: Build minimal viable unit on OpenSearch following Analyzer → Index Model → Query Strategy, replay real traffic for verification.

Performance Tuning & Linear Scaling Test: Iterate “resource config → query rewrite → index split” three times, then conduct linear load testing after scaling.

Canary Release & Traffic Switching: Gradual traffic migration by proportion or business dimension, with real‑time monitoring to ensure data consistency before full cutover.

Operations Harden & Ongoing Management: Deploy time‑based scaling and deletion protection, set up dashboards for latency, error rate, and resource utilization.

Conclusion & Outlook

The GoTerra case validates Alibaba Cloud OpenSearch industry‑algorithm edition’s capabilities in multilingual semantic understanding, hybrid vector retrieval, extreme performance, and elastic cost control. It provides a reproducible, end‑to‑end intelligent search framework that can be generalized across enterprises, while the AI Search team continues to push core innovations in cognitive intelligence and cross‑modal retrieval.