How DoorDash scaled merchant-facing analytics to thousands of users in weeks, on dbt and ThoughtSpot

https://www.youtube.com/watch?v=FM_sFQsKUw4

In this fireside chat, Sonny Rivera, principal data and AI strategist at ThoughtSpot, interviewed Harsha Reddy, senior engineering leader at DoorDash, who leads Data End-User Platforms within DoorDash's data platform organization, about the company's path from prototype to production analytics.

Harsha framed DoorDash's core data strategy as balancing two goals that are often in tension: central governance and user velocity. That tension is compounded by DoorDash's acquisition history. Deliveroo (UK) and Wolt (Finland, and much of the EU) each arrived with their own semantic layers, often built in Looker, on top of DoorDash's own in-house semantic tooling.

Consolidating those into one governed layer, without slowing teams down, became the actual problem to solve. DoorDash's approach splits the semantic layer from the consumption layer. Data lands in a data lake (exposed as Iceberg tables) and in Snowflake, dbt defines the semantic and metric layer on top of both, and ThoughtSpot is one consumption layer among others, connecting directly to dbt's metric definitions rather than requiring its own parallel set.

As Harsha put it, that's the advantage over an all-in-one tool like Looker, where "there's just one tool where it's got everything encoded, and it's not so much useful outside for anyone else."

That structure supports two very different use cases today. Internally, for finance and revenue reporting, where Harsha said "the margin for error is zero," the governed semantic layer combined with ThoughtSpot's built-in compliance and audit reporting replaced work DoorDash used to do by hand to satisfy GDPR, HIPAA, and other regulatory requirements.

Externally, DoorDash embeds ThoughtSpot, fully white-labeled with no ThoughtSpot branding visible, directly into merchant-facing tools, with role-level access control ensuring one merchant never sees another's data. Harsha declined to share an exact figure but said DoorDash went from a new use case to "thousands of users within a few weeks."

Asked for advice, Harsha's answer was direct: semantic modeling is hard, but in a world where generating an answer is cheap and trusting it is what's valuable, it's worth the effort. Treat the semantic layer as code (version-controlled and reviewed), and make the governed path at least as easy as writing an ad hoc query, or people will route around it.

On where AI fits in, DoorDash has moved from a set of domain-specific agents behind a router to a single global agent grounded in the same dbt-defined semantic layer that feeds its BI tools, a model that now covers roughly 70 to 80 percent of DoorDash's data domains.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →