4 Tips for Shipping Data Products Fast

Summary

Shipping data products under tight deadlines requires teams to stay focused while avoiding overengineering and building something users will not use. Shopify’s Data Science team recommends four practices: run a timeboxed design sprint with a small team and a designated Champion, prototype quickly—often with spreadsheets—and gather feedback throughout development. For first iterations, the post advises using simple approaches instead of machine learning when complexity would delay integration, polish, and user conversations; a sales forecasting prototype used linear regression with two input variables before a more complex model was built. User research should use specific questions and a diverse set of users, as feedback changed a confusing 0–100 scoring display to easier-to-interpret letter grades. Together, these practices provide a clearer path to launching useful products quickly.

Context

Shopify’s Data Science team needs to build and ship data products quickly because commerce changes rapidly and its products support merchants throughout their journeys. Tight timelines create risks including endless design phases, overengineered machine-learning systems, and products that users do not use.

Approach / What changed

The post recommends timeboxed design sprints with small teams and a Champion, rapid prototyping with tools such as spreadsheets, simple non-machine-learning approaches for first iterations, and structured user feedback gathered before, during, and after each build iteration.

Takeaways

• A design sprint typically lasts up to five days, uses a team of no more than six or seven people, and assigns a Champion to drive direction and make decisions.
• A prototype can combine SQL queries, a bash script, and spreadsheet formatting; Shopify used this approach for a marketing attribution tool to make changes and comparisons quickly.
• Feedback on a sales forecasting tool showed that numerical scores from 0–100 were confusing, while letter grades such as A, B, and C were quicker for users to interpret.

Related reading