Zappos Creates Breakthrough Customer Experiences Using AWS

Building a trusted ecommerce brand requires a focus on customer experience innovation. Especially in the apparel segment, traditional, hands-on retail experience has shaped customer expectations. In the words of Ameen Kazerouni, head of machine learning research and platforms at Zappos, “The holy grail of ecommerce is to get customers to feel as comfortable as they would in a store—knowing how an item fits, how it looks, with as much confidence as if they were physically trying it on.”

Achieving this type of personal experience in a digital store is no small task. But if any company is up to it, it’s Zappos. Established in 1999, Zappos.com is a major online apparel retailer famous for its industry-leading, 365-day return policy, fast and free shipping, and 24/7 service. Since 2009, Zappos has been a subsidiary of Amazon.

The company measurably improved the ecommerce customer experience using analytics and machine learning built on Amazon Web Services (AWS). The solution enables Zappos to personalize sizing and search results for individual users while preserving a highly fluid and responsive user experience.

Zappos knows that providing accurate recommendations is key to an efficient shopping experience. The company reassures customers with its generous return policy and fast and free shipping, but these offerings are both expensive and undifferentiated.

“We are always asking ourselves: how do we differentiate further?” says Kazerouni. “How do we optimize return rates without negatively affecting the customer experience? These are the problems we set out to solve using machine learning and analytics on AWS.”

In the search phase of the customer journey, the company’s goal was to make personalized recommendations at runtime to increase search relevance. Rather than using a generic search algorithm, Zappos seeks to understand customers personally and provide a unique set of search results for a given term. (It also prominently displays an opt-out button for customers who do not want this level of personalization.)

At the same time, it can’t afford to slow down search performance noticeably. “We needed to minimize the amount of time the extra operations take,” notes Kazerouni. “So we combine high-performance caching, strategic precalculation of certain results, and ensemble-based machine learning approaches that use multiple, simple models.”

The data pipeline starts with a lightweight client sending relevant events to an ingestion API for processing. The API sits in an auto-scaling group to handle high volumes of data. From the API, the data is sent to Amazon Data Firehose for ingestion into an Amazon Redshift data warehouse that provides high-performance data access for machine learning research. Amazon Simple Storage Service (Amazon S3) is the intermediary between Amazon Data Firehose and Amazon Redshift.

Zappos uses several technologies for training and running models. It relies on Amazon SageMaker to predict customer apparel sizes. These predictions are cached and then exposed at runtime via microservices APIs for use in recommendations. Zappos uses Amazon EMR to run big data analytics for a fraction of the cost of traditional on-premises clusters. It also runs models using graphical processing units (GPUs) on Amazon Elastic Compute Cloud (Amazon EC2).

The company enables ultrafast lookup of precomputed predictions using two distinct services. Amazon DynamoDB stores precomputed results that will be accessed at runtime. This fully managed key-value and document database delivers single-digit millisecond performance at almost any scale. It can handle more than 10 trillion requests a day and can support peaks of more than 20 million requests per second. For even faster response times, Zappos takes advantage of Amazon ElastiCache for Redis, an in-memory data store, as a cache layer. This service ensures sub-millisecond latency when needed.

The microservices that run models and consolidate results run on Amazon EC2 instances arranged in auto-scaling groups with location-based load balancers. Zappos uses Amazon Route 53 as the domain name system, routing traffic throughout the solution.

Creating and maintaining this intricate architecture with traditional development and deployment methods would be prohibitively complex. Instead, Zappos relies on infrastructure as code using AWS CloudFormation. “Every aspect of the solution is represented in AWS CloudFormation templates,” reports Kazerouni. “To make a change, we just tweak the template. If we need to fix the way the services communicate with Redis, we don’t repeat the change manually—we change the template and deploy it everywhere.”

He notes that it would be impossible to build the solution without the wealth of AWS services at the team’s command. “Using AWS services as building blocks allows engineers to focus on improving performance and results rather than DevOps overhead.”

Zappos delivered these improved search results to customers with a nearly undetectable increase in latency, with 99 percent of searches completed in less than 48 milliseconds. By using a similar architecture, it has also significantly improved personalized sizing recommendations based on simple fit surveys and past purchases. As a result, the company has reduced repeated searches and product returns. It has also achieved higher search-to-product-clickthrough rates and raised the position of customer selections in search results.

As Kazerouni sums it up, “We think of ourselves as a customer service company that happens to sell shoes and apparel. Anything we can do that improves service improves our business. Using AWS makes it possible for us to innovate the experience faster.”

To learn more, visit aws.amazon.com/big-data/datalakes-and-analytics.