Depop Case Study

Depop offers an alternative to shopping through its marketplace for unique fashion. It turned to AWS after reaching the limits of its existing PaaS technology. AWS enables Depop to quickly iterate, deploy, and scale out new features, with a focus on machine learning.

Depop bills itself as a community-driven social shopping experience where the next generation buys, sells, and discovers unique fashion. Depop aims to enable the fashion industry to become more sustainable and purpose-driven by offering a circular-first alternative to shopping new.

To support growth and the new capabilities it would need in the future, Depop went all-in on AWS. It has since evolved from an early-stage startup to a late-stage success story with 25 million users (90 percent of its active users are under the age of 25) and year-on-year customer growth of 130 percent across its key markets of the US, UK, and Australia.

Depop moved to AWS around two-and-a-half years ago, as it was hitting the limits of its existing Platform-as-a-Service (PaaS) technology. “With our PaaS provider, we were facing several challenges, such as scalability limits, lack of flexibility in how we scale our applications, and constraints in the way solutions are built due to a fully managed, but limited, set of available capabilities,” says Depop chief technology and data officer Remo Gettini.

Robert Erdin, Depop director of applications and services, adds that the company also needed a more cost-effective infrastructure over which it had more control.

Depop performed the migration itself, with AWS solution architects helping to overcome challenges. Those challenges included understanding the various cost-saving measures and how to consistently monitor and apply them, as well as the maturity of different AWS services.

Another challenge was finding the right balance between providing reusable, secure, and easy-to-use abstractions for application developers, and granting developers direct access to provisioning infrastructure.

Thanks to using AWS, the company has increased the number of teams able to simultaneously work on the Depop application. That number has grown from two to eight, with more than 10 teams possible soon. This has greatly improved Depop’s ability to develop, test, and deploy new services. Clemence J. Burnichon, Depop head of data science and machine learning (ML), says AWS has also given her team the flexibility to expand from a two-instance cluster to 25 instances.

AWS also supports Depop’s rapid market growth. The Amazon CloudFront content delivery network, integrated with Amazon Simple Storage Service (Amazon S3), will be crucial to facilitating this expansion by supporting the distribution of images and videos.

Depop’s services are increasingly driven by ML with the company making extensive use of the AWS ML service for its ML workloads. One of the main technology challenges for Depop is dealing with an ever-growing inventory in which no two items are the same. The company relies on the Amazon S3-based data lake solution to manage its vast inventory of 25 million items and transactions, employing Amazon Kinesis Data Firehose to stream data, with some use of Amazon Managed Streaming for Apache Kafka (Amazon MSK).

Combined with visual recognition technology, ML algorithms categorize clothing items held in the data lake in different ways, including by size, color, and brand, supporting Depop’s Personal Shopper search and recommendation service.

Depop plans to use Amazon Elasticsearch Service on its mobile app to take advantage of the categorized data held in the data lake. It will provide more granular control to search and enable more ML algorithms to be added over time.

Other Amazon technologies used in conjunction with the data lake include the Amazon Redshift data warehouse for creating cleaner, packaged versions of data, and the Amazon Athena interactive query service to provide rapid access to data.

As a result of having these AWS tools at its disposal, Depop’s ML team is also able to quickly iterate new deep-learning models. According to Burnichon, the ML team currently has around 30 ML models in production, processing a combined 1.5 million messages per hour.

These models are supported by Amazon SageMaker and Amazon EMR for indexing and relevance, while extract, transform, and load (ETL) processes are provided via AWS Glue.

AWS has allowed Depop to move toward “more granular backend services, which allowed us to significantly scale up our engineering team and work on more features concurrently,” explains Gettini.

Depop is using Amazon Elastic Kubernetes Service (Amazon EKS) to provide serverless compute for containers that support its mobile app, giving the development team the ability to deploy new production services within a day. Erdin says AWS has enabled a “democratization of access to infrastructure,” allowing the development teams to easily and cost-effectively try out new things in a safe way, without incurring technical debt. “No matter what you want to do, there is an AWS service for it,” he says.

Using AWS also ensures that successful services or applications are already in the right environment to go into production. According to Erdin, Depop has seen more than 100 services go live.

Another major benefit is the access that Depop has to AWS specialists. Erdin cites the example of being able to unblock ideas by bouncing questions off of AWS solution architects. AWS also provides Depop with frequent training, from entry-level sessions about AWS and cloud to deep dives into specific technologies.

For Gettini, Depop’s biggest achievement with AWS is being able “to support a product and engineering organization of over 100 people in nine cross-functional teams with only a handful of engineers dedicated to maintaining infrastructure.”

“Comparing this to the beginning of my career almost 30 years ago, it is still mind-blowing.”