Evolving Search Architecture for a Demanding Tesco Online Business by Anand Jayaram

Summary

The presentation covers the evolution of search architecture at Tesco, focusing on improving search relevancy and scalability for a growing online business. Key strategies include data liveliness, leveraging Kubernetes and containerization for flexibility, and implementing vector search algorithms for more accurate and context-driven results. The talk highlights the challenges of creating a personalized, efficient search experience amid intense competition and rising customer expectations. It also emphasizes the importance of simplicity, cross-team collaboration, and cost-consciousness in building a scalable, future-proof architecture that can serve various internal stakeholders and customers.

Takeaways

• 😀 Evolving search architecture for online businesses requires balancing art and science due to the complex nature of signals involved in search algorithms.
• 😀 Personalization is a crucial aspect of search, considering factors like product popularity, seasonal promotions, and customer preferences.
• 😀 The increasing revenue opportunities through search make it essential to adapt and scale search technologies to meet growing customer demands.
• 😀 Key principles for improving search include data liveliness (keeping data up to date), scalability, user experience, cost efficiency, and enabling other teams to benefit from the infrastructure.
• 😀 A significant upgrade to the search engine (Elastic 8.x) resulted in 30% reduction in storage and improved query performance, despite the complexity of the upgrade process.
• 😀 Kubernetes was adopted to improve the scalability and flexibility of the search infrastructure, allowing for seamless workload portability between cloud providers.
• 😀 By breaking down VMs into containers and using Kubernetes, the company optimized its licensing costs with Elastic while achieving better resiliency.
• 😀 Vector search, a hot topic in search, enhances user experience by enabling more relevant search results based on user intent rather than just keywords.
• 😀 The architecture includes a feeder, enricher, and indexer to efficiently process data, making it available for real-time search while also handling business-specific customizations for different regions.
• 😀 Kafka and its stream processing were key in achieving real-time data integration, with a streamlined architecture that scales and supports other teams within the organization.
• 😀 The adoption of simplicity in architecture and constant monitoring through tools like New Relic helped ensure system reliability and effective risk management in production environments.

Q & A

What is the primary goal of evolving search architecture for Tesco?

The main goal is to meet growing business demands by improving search relevance, personalization, and scalability. The focus is on ensuring data liveliness, enhancing user experience, and enabling efficient cost management.

What challenges did Tesco face with the current search engine before the upgrade?

The challenges included handling an overwhelming number of search signals, such as sponsored content, seasonal promotions, product popularity, and personalization, all of which had to be balanced for an optimal search experience.

How does search differ from other processes like payments, according to the speaker?

Search is not deterministic like payments, meaning that it is based on heuristics and requires a combination of art and science to fine-tune various signals, while payments follow a straightforward, predictable process.

What are the key business drivers for enhancing search at Tesco?

The key drivers include changing customer expectations, growing revenue opportunities, expanding product ranges, fierce competition, and the need for cost-efficient scalability.

What were the guiding principles behind the architectural overhaul of Tesco's search system?

The principles included ensuring data liveliness, scaling the system efficiently without excessive costs, providing a store-like experience for users, and enabling collaboration across teams to ensure that everyone could benefit from the architecture.

What specific changes were made to improve the search engine's performance?

The search engine underwent an upgrade to Elastic 8.x, which improved query performance and reduced storage needs by 30%. Additionally, the system transitioned from virtual machines to Kubernetes with containers, enhancing flexibility and scalability.

What role does Vector Search play in Tesco's search architecture?

Vector Search allows for capturing user intent more accurately, especially in cases of synonyms or related terms. For example, it helps understand that 'soccer' in the US means 'football' elsewhere, enabling more relevant search results.

How does Tesco utilize Kafka and stream processing in its architecture?

Kafka is used as a central messaging system to ingest and process data in real-time. The architecture includes components like feeders and enrichers that handle event-driven data and ensure it is indexed correctly for search.

What benefits were realized from implementing vector search in production?

Vector search improved search relevance significantly. For example, when searching for 'Kit snacks', vector search returned more appropriate results like strawberry fruit and jellies, as opposed to irrelevant results found through traditional keyword search.

What lessons did Tesco learn during the implementation of the new search architecture?

Key lessons included the importance of simplicity, aligning technical decisions with business needs, and the necessity of ensuring backward compatibility. The implementation also highlighted the importance of gradual, risk-managed changes rather than big bang releases.