May 2022 Summaries

Uber's reliance on operational machine learning has been key to its success, with the company leveraging this approach to power real-time decision making that directly impacts the end-user experience. Operational machine learning is distinct from analytical machine learning, focusing on autonomous and continuous decisions in production environments. The trend of modernizing data architectures and adopting DevOps principles has made it possible for companies to adopt operational ML, enabling rapid iteration and deployment. To get started with operational machine learning, businesses should identify suitable use cases, empower small teams, prioritize high-impact applications, and learn from others' experiences. With the right approach, operational ML can be a key differentiator between winners and losers in various industries.

The article highlights the challenges faced by central data teams in organizations that use machine learning, and how these teams often fail due to a lack of integration with the business. The author argues that embedding data specialists in product teams is a more successful approach, allowing for better collaboration, improved product planning, and increased opportunities for machine learning adoption. This approach also leads to clear ownership and accountability, as well as the ability to identify and address performance issues earlier on. By contrast, central data teams often struggle with integrating models into the business, managing data pipelines, and addressing performance degradation over time. The article concludes that while there are challenges associated with embedded teams, the benefits far outweigh the costs, making this approach a more successful strategy for organizations looking to leverage machine learning effectively.

The text discusses the importance of feature stores extending, rather than replacing, existing data infrastructure. It highlights the limitations of isolated ML project stacks and the benefits of integrating feature stores with underlying data platforms to centralize and make available ML data. The author emphasizes that ML has unique requirements, but also shares common non-unique requirements, which should be addressed by optimal ML data infrastructure design. Feature stores aim to maximize the use of existing platform and infrastructure while supporting the best workflows for AI/ML developers. The text concludes with a call to attend Apply(conf), a main event for the year, featuring speakers working on fascinating work in ML data engineering.

Emma Peng is a Software Engineer at Tecton, where she works on online storage features such as data freshness and low-latency lookups for customers using DynamoDB and Redis. She joined Tecton from a startup called Coolan, which was later acquired by Salesforce, after working in various roles including monitoring and analyzing hardware infrastructure and building healthcare case management systems. Emma values the sense of achievement and autonomy that comes with working at a startup, where she can contribute to high-impact projects and deliver without red tape. She finds it gratifying to explore different interests with her coworkers, who have diverse hobbies and passions outside of work, and appreciates the comfortable Tecton swag, including an autographed football and a Lululemon crewneck sweater. In her free time, Emma enjoys taking photos, capturing people's emotions, and using Google Photos to relive fun memories.