Brief
The essential idea
Uber operates global, real-time Mobility and Delivery products with more than 4,000 engineers, thousands of services, and hundreds of thousands of deployments per week. Since 2023, the organization has distinguished Program leadership, which drives cross-functional outcomes, from Platform leadership, which builds reusable technical capabilities for many programs.
The architecture evolved from a monolith to microservices around 2016, then from a largely in-house infrastructure stack toward a multicloud model with Google Cloud and Oracle announced in 2023. Up replaces the older microservice deployment system with a portable workflow that can target both internal and cloud environments, including standardized Arm profiles.
Uber's ML platform evolved in waves: predictive ML and Michelangelo in 2016-2019, deep learning in 2019-2023, and generative AI with a central gateway and LLMOps after 2023. The common thread is platformization: repeated infrastructure and governance problems become paved roads so product teams can focus on domain outcomes.
Decision lens
Key takeaways
Program leaders coordinate outcomes across domains, while platform leaders create reusable leverage across programs.
Microservices solve some scaling problems but require strong deployment, ownership, and observability platforms.
A portable deployment abstraction reduces coupling to a single infrastructure environment.
Multicloud adoption should preserve one developer workflow rather than exposing every provider difference to teams.
ML capabilities mature from individual models into shared training, serving, governance, and LLMOps platforms.
Platform investments should be justified by reduced cognitive load and faster product delivery.
Workplace experiment
Apply it at work
- 1
Separate one cross-company outcome into program responsibilities and reusable platform responsibilities.
- 2
List infrastructure assumptions embedded in the current deployment workflow and identify which should become portable abstractions.
- 3
Measure whether a paved road reduces lead time and operational burden for at least two product teams.
- 4
Define one governed entry point for generative-AI models, including access, evaluation, cost, and observability.
Choose one action, define the observable effect, and keep the first test small enough to reverse.
Evidence
Sources and further reading
Additional sources
Channel, aggregator, and commentary links confirm the work; they are not the primary source.
Local knowledge map
A small, typed neighborhood instead of the full-catalog graph.